OpenAI framed ChatGPT’s Instant Checkout as obvious: more than 700 million people use ChatGPT each week for everyday tasks, many of them shopping queries. Now, the assistant can help you buy directly. Powered by the Agentic Commerce Protocol (ACP), co-developed with Stripe, the feature debuts with U.S. Etsy sellers and soon over a million Shopify merchants, from Glossier to SKIMS.

The pitch is user simplicity: describe a product, see the best matches, tap to buy without leaving chat. For merchants, the promise is reach and relevance, with ACP making integration lightweight. For users, the experience is free, secure, and ranked purely on relevance, not fees.

Many conversational shopping assistants, comparison bots, plugin SDKs, and browser tools have become endangered species within the last 24 hours. The moment OpenAI can go from “help me find” to “help me buy” inside chat is a major shift in funnel control – those that don’t pivot quickly get squeezed.

And will retailers and brands quickly adapt to the new reality of yet-another-channel in a multichannel world? Or will they forget the lessons of mobile web and app adoption and take years to catch up with consumer preferences?

In vitro

2024–25 studies demonstrated that conversational AI is not just feasible but advantageous in e-commerce.

• Graph-enhanced retrieval raised factual accuracy by 23% and improved first-contact resolution by 28%.

• LLM-based assistants like Flippi reached >90% accuracy across 1.3m users, with measurable conversion uplift.

• User research shows AI-driven shopping doubles repeat purchase intent by reducing decision fatigue.

In vivo

The market followed: conversational AI commerce is projected at $290B for 2025, with generative AI referrals to retail sites up 4,700% year-on-year. Instant Checkout brings those findings to market scale. Key moves include:

• Coverage: Etsy live today, Shopify’s million-plus merchants next.

• Capability roadmap: single-item purchases now; multi-item carts, more geographies, and more processors coming.

• Open standard: ACP is open-sourced, designed to work across processors via Stripe’s Shared Payment Token API or ACP’s Delegated Payments Spec.

• Merchant control: sellers remain merchants of record; they handle payments, fulfilment, returns, and support.

• Trust guarantees: encrypted tokens, minimal data sharing, explicit user confirmation on every step.

The effect is a closed loop: discovery, recommendation, and purchase inside the assistant, with governance claims embedded at protocol level.

Trojan horse

Behind the convenience is strategic leverage. To compete in this new channel, merchants must structure their data, expose APIs, and otherwise raise their AI game. In effect, OpenAI is inducing the retail ecosystem to modernise.

This dovetails with NVIDIA’s ambition: every upgrade to personalisation, inventory prediction, or recommendation is another GPU workload. ChatGPT’s checkout feature is thus a Trojan Horse: pulling enterprises into deeper AI adoption under the banner of consumer convenience.

From human-in-loop to A2A

Looking out to the near future we can see a 3 step strategy unfolding:

1. Human-in-loop commerce where users confirm each purchase.

2. Policy-guided automation for example “auto-reorder under $30,” “book flights under $500.”

3. Full agent-to-agent commerce when assistants transact directly with merchant agents under policy and audit, not per-click consent.

ACP is architected for all three. Today it enforces human confirmation. Tomorrow it enables autonomous negotiation between agents.

This isn’t just speculation. Early frameworks are emerging that attempt to map what a genuine agent-to-agent economy looks like, distinguishing today’s AI-assisted checkout from fully autonomous A2A transactions. Work from Stelia points to the need for a holistic approach: defining taxonomies of agents, building economic and legal architectures, and stress-testing failure modes with scenario modeling. The research agenda spans technical feasibility, market mechanisms, governance, and social impacts out to 2035. In that context, OpenAI’s Instant Checkout looks less like a convenience feature and more like the first policy-constrained implementation on a much longer trajectory.

Platform gravity

As shopping, payments, and conversation converge, workflow gravity increases. Just as GPT-5’s tool use will increasingly pull SaaS workflows inside the model, commerce flows will settle in ChatGPT. Control shifts to the platform that captures intent and closes the sale in a single chat window.

Safety that travels well

OpenAI’s PR emphasises trust: minimal data sharing, explicit confirmation, secure tokens. But as ACP scales to A2A, safety must expand: protocol-level constraints on agent behavior, auditable logs, and jurisdiction-aware compliance. Without verifiable safety, merchants and regulators will resist deeper automation.

Orchestration

Global vertical orchestration platforms such as Stelia are becoming the benchmark substrate beneath this shift. They orchestrate “ground to prompt”: racks, network kernels, runtimes, and interfaces. In agentic commerce, these platforms guarantee:

• Latency low enough for real-time checkout.

• Compliance with payment and data laws across regions.

• Resilience across multiple processors.

• Auditability from GPU kernel to consumer receipt.

They will be the execution fabric for agentic commerce at planetary scale.

Strategic, not opportunistic

On the surface, Instant Checkout is a new feature and revenue stream. In substance, it is a deliberate long-term bet:

• Capture intent at the source.

• Normalise ACP as the open transaction rail.

• Pressure merchants to upgrade their AI maturity.

• Indirectly advance NVIDIA’s goal of AI saturation across industries.

Convenience is the Trojan horse. Control of the rails is the strategy.

By Paul Morrison, Chief Experience Officer, Stelia

Subscribe now

Sam Altman opened the recent launch with a simple arc. Whilst GPT-3 felt like a bright sixth former and GPT-4o like a capable undergraduate. GPT-5, he said, behaves like a PhD-level expert on demand. He also said 700 million people now use ChatGPT each week. That scale changes the product story. It also changes the governance story.

What follows looks from the Stelia perspective at what gets built, who holds leverage, and which architectural choices age well when intelligence starts to act through tools as much as words.

SaaS – canary in the coal mine

For two decades, SaaS companies froze workflows into persistent products. GPT-5 shows a different rhythm. In the launch demos, an agent turned a plain-language request into an interactive aerodynamics simulator in minutes. It produced hundreds of lines of running code with no IDE in sight. The SaaS example illustrates how a fixed product can be replaced by a temporary configuration that vanishes once the job is done. However, the more important shift is the growth of tool use.

GPT5 edge: autonomous toolchains

GPT-5 is producing even more human-like text, yet it is inow also initiating multi-step processes.

• Opening files, parsing and rewriting them.

• Calling APIs with structured parameters.

• Searching, filtering and reasoning across results.

• Handing output to other tools without manual intervention.

The demos showed it working for minutes at a time, chaining tools and functions into coherent workflows. It can run the equivalent of an internal build system, a customer research desk, or a legal discovery process. All from a prompt. This level of autonomy moves well beyond the earlier “search and summarise” agents.

Strategically, the point of control has shifted to the interface between the model and its tools. That is the new coordination substrate.

Brand in a tool-driven world

When models and connected tools operate other systems, brand becomes a claim about execution integrity.

Enterprises will ask:

1. Will this agent only call approved tools, in the way I approved?

2. Will it respect the jurisdictional and contractual constraints I operate under?

3. Can I verify and audit what it did, step by step?

Brand in this context is built on governance primitives. These include policy as code, verifiable attestations for every tool call, and cultural-intelligence models that adapt without human babysitting.

Platform gravity

GPT-5’s integration with Gmail, Calendar, Python data analysis, and image generation widens the set of jobs that can be completed without leaving the platform. When tools, memory, and history all live together, workflows settle in and stay. Controlling that space is controlling the habit, hence the near-daily cadence of model upgrades and habit-forming feature enhancements.

Safety that travels well

Once a model can act through tools, the potential blast radius grows. Guardrails must move from output filtering to protocol-level constraints on tool invocation and data flow. Post-hoc review helps but cannot cover long-running or multi-system processes.

Safety here means:

• Pre-deployment ethics checks on all default tools.

• Context-aware limits on tool combinations.

• Transparent logging so actions can be replayed and audited.

The coordination challenge is making these properties travel with the task, across every tool it touches.

Costs, compute

Like increasingly complex SaaS led to inevitably opaque billing, tool-driven workflows are not free. Each API call, file parse, and image render adds latency and cost. The platforms that endure will be the ones that manage predictable economics without hollowing out control.

That points to three design moves:

1. Structure reasoning so outputs are easy to constrain and validate.

2. Shift suitable steps to the edge to cut round trips.

3. Cache where privacy allows.

Planet-scale coordination

As tool use scales, so does the need for responsible coordination. A scheduling agent will email your client. A coding agent will merge to main. A finance

agent will touch a ledger. These scenarios are already in use.

The substrate, such as Stelia, keeping these agents at a predictable cost, lawful, safe, and interoperable across borders is not optional. It is the foundation for an intelligence layer that can operate anywhere without eroding human dignity or autonomy.

Subscribe now

In my last article, I explored the orchestration challenges facing modern media companies. We looked at how global workflows and fragmented toolchains are forcing publishers and broadcasters to rethink how they produce, localise, and distribute content at scale. Now, as I meet with partners and clients across the entertainment industry in New York, through meetings, studio visits, and events like the Monks happy hour, I am seeing the shift move further. What began as media transformation is rapidly becoming a full-scale restructuring of entertainment. And this time, the change is not coming from inside the studio walls. It is being shaped at the system level.

The world of entertainment used to be studio driven, rooted in the comfort of predictable, linear consumption. Now, it is being driven by behaviour, emotion and cultural dynamics that change by the hour. The old question: what are we producing is being replaced by a more urgent one: how are we adapting what we produce to real-time attention? That question is reshaping production cycles, marketing strategy, rights management, and even how executives evaluate ROI. The infrastructure of entertainment, not just the content itself, is being rebuilt.

Netflix creative scale as competitive edge

A recent case from Netflix highlights how different that infrastructure now looks. In their new Argentinian sci-fi series The Eternaut, Netflix used AI-powered VFX through Eyeline Studios to create a building collapse scene that would have been prohibitively expensive to film traditionally. According to Co-CEO Ted Sarandos, the sequence was delivered ten times faster and at a significantly lower cost. The results speak for themselves. In the same quarter, Netflix posted over 11 billion dollars in revenue, up 16% year over year. More than visual effects, this lesson is about how infrastructure enables creative scale. Netflix’s investment in real-time, AI-enhanced tools is allowing them to tell stories they otherwise could not afford and to do so with local teams producing globally resonant results. This infrastructure-first approach to creative scale is now spreading across the industry, but it requires new ways of measuring what actually works.

What MediaScience taught me about emotion

Understanding what works means measuring real human response. Last week I visited MediaScience’s lab in Manhattan, where they test content by measuring real human response eye movement, facial expression, electrodermal activity, and heart rate. Their research shows that attention is shaped in the first seconds and influenced by emotion, context, and platform. I watched how small creative choices – a change in music, a tighter edit – triggered measurable shifts in engagement. Studios are already using it to green-light projects, fine-tune trailers, and adjust platform recommendations. MediaScience found that adapting stories based on behavioural signals can boost viewer engagement by over 30%.

What struck me most is how context and psychophysiological data has become a strategic input. Rather than focusing solely on views or completion, the emphasis has shifted towards understanding and quantifying emotional context. MediaScience is measuring what the viewer feels, right now, and how content responds. This is the new layer of intelligence shaping how entertainment is made and delivered. Phillip Lomax, MediaScience EVP, explained that context is shaped by four pillars, Audiences, Platforms, Content, and Ad Formats, and how MediaScience has gathered significant intelligence across these subjects which have been leveraged by their clients at Disney, Comcast, Google, Amazon, Netflix and many others.

Why platforms like Tubi are winning

That emotional intelligence MediaScience measures is exactly what’s driving the fastest-growing platforms. During my meetings here in New York, it has become obvious that many of these platforms are not focused on subscribers alone. They are building ecosystems around engagement. Tubi, for example, recently surpassed 100 million monthly active users, with over a billion hours of content watched in a single month. These are ad-supported businesses with no subscription revenue. Instead of payment, their metric of success is attention, relevance and repeat engagement. And they are winning by making their AI-enabled infrastructure flexible enough to adapt to viewer behaviour in real time. That adaptability is increasingly powered by AI systems that interpret how, when, and why people engage.

In this new wave of entertainment, the real competition isn’t just between streamers, it’s between formats. Premium, paid experiences on one side; free, algorithm-driven discovery on the other. And the future likely sits somewhere in between. The next generation of platforms is betting on frictionless, culturally-tuned content that feels as easy and diverse as scrolling TikTok, while still delivering the narrative pull of long-form entertainment. There’s also a shift in mindset: from producing more to producing what connects. Experiments are becoming institutional like studios that green-light based on fan input, or libraries that blend creator-led stories with licensed hits. What’s emerging is a new hybrid model where the old Hollywood playbook is rewritten for a leaner, faster, more emotionally aware viewer economy.

Studios focus on convergence, not channels

That same mindset is shaping strategy at the studio level. With linear viewership down and theatrical releases under pressure, the big players are thinking convergence. Disney is in the final stages of consolidating its operations across Hulu, ESPN, and Disney+ into a single unified offering making Disney+ its one-stop shop in streaming. The same time Netflix is investing heavily in music with reports of deeper partnerships spanning video, music, and live events. The experience is now the differentiator. And experiences require infrastructure that can handle rights, regionalisation, creative variation, billing, and measurement all in one real-time loop.

Fox moves toward one spine

FOX is also embracing both ends of the format spectrum. On one side is Tubi, their ad-supported platform built for scale and discovery. On the other is FOX One, a premium, subscription-based service launching August 21. Meanwhile, their internal OneFOX initiative unifies content, advertising, and metadata into a single infrastructure layer. When companies of this size start operating from a single spine, it changes not just internal speed. It changes what becomes possible for audiences and advertisers alike. Together, these moves reflect FOX’s twin approach to market: accessible at scale, integrated at depth.

When AI orchestration works

Yet most companies remain stuck in what I call the pilot trap. They are experimenting with artificial intelligence tools here and there, layering plugins or point solutions into isolated workflows. But they are not integrating. They are not feeding data back into their systems. And as a result, they are not scaling. True orchestration does not come from tool trials. It comes from connected systems that allow real-time response to audience behaviour, creative feedback loops, and operational decision-making.

You can see what this looks like when it works. In June, Netflix released the animated film KPop Demon Hunters. The movie was watched over one hundred thirty million times within weeks, and its soundtrack dominated the charts across the United States and Asia. Fan art began circulating within days. Cosplay, TikTok remixes, and micro-fandoms exploded globally. What made the real difference was the way AI models tracked and adapted in real time to audience engagement, shaping marketing and creative simultaneously. Much of that responsiveness came from AI models tracking sentiment and engagement signals across platforms.

From creative execution to strategic infrastructure

These successes are forcing teams across the industry to ask serious questions. How do we adapt narratives across cultures without reshooting entire scenes? How do we automate marketing assets in multiple languages within hours? How do we ensure our rights logic and licensing structures do not become bottlenecks to creative agility? These strategic questions are the ones shaping who moves forward and who gets left behind.

What the Paramount-Skydance deal reveals

The newly sealed merger between Paramount and Skydance confirms precisely this shift. It brings together Hollywood legacy and Silicon Valley agility, creating a powerhouse built for an AI-shaped era where flexibility and infrastructure matter as much as storytelling. Editorial integrity, platform integration, and operational scalability are now board-level priorities. For the industry, it indicates a new phase - one where collaboration is creative and systemic.

Programming for cultural moments

During one of my conversations this week, an executive said something that stuck with me. He said, we are no longer in the business of programming for time slots. We are programming for cultural moments. That line captures exactly where entertainment is heading. The winners will not be those who simply produce more content. It will be those who can connect that content to emotion, context, and culture – fast, at scale, and across borders.

Whether it's a media agency like Havas aligning production and performance, or a contextual tech company like GumGum measuring attention with AI-powered contextual signals, the lines between storytelling, infrastructure, and intelligence are blurring fast.

Entertainment’s new operating system

Across all of these examples from Netflix’s AI-powered production to Tubi’s adaptive engagement engine and GumGum’s contextual intelligence, a new layer of AI infrastructure is quietly becoming the foundation of modern entertainment. It’s not just content being transformed, but the entire operating system behind it: one powered by intelligent agents, audience signals, and dynamic orchestration.

I am seeing this first hand across every conversation I am having now in New York. Whether I am speaking with creative studios, rights managers, marketing leaders, or platform heads, the same trend keeps emerging. The entertainment industry is undergoing a complete restructure. The systems behind the scenes, the tools, workflows, orchestration engines, and attention models, are now what make or break success.

It is no longer enough to have good content. You need good systems. You need the ability to operate across markets, adapt in real time, personalise without compromising quality, and measure value not just in views but in cultural impact. The companies that build this AI-native infrastructure now will not only lead in entertainment. They will define how entertainment itself is experienced in the years to come.

by Ula Nairne, VP Media & Entertainment, Stelia

Subscribe now

Climate disaster costs

In 2024, weather‑driven catastrophes wiped out US $320 billion in global wealth, more than the GDP of Greece, with floods and wildfires leading the tally.¹ Insured losses have risen about 7 % each year since 2015, outpacing world GDP. The planet is paying a premium for information it still receives too late.

Defences falling behind

Warmer oceans, erratic rainfall and expanding wild‑land-urban interfaces have doubled flood‑ and fire‑related losses over the past decade. Policymakers are scrambling: the UN‑backed Early Warnings for All initiative pledges universal, life‑saving alerts by the end of 2027.² Capital is flowing the same way - venture funding for AI‑powered Earth‑observation (EO) start‑ups jumped again in 2024 because early, trusted intelligence is fast becoming critical infrastructure.

Satellite breakthroughs

Why only AI can keep up

EO satellites now pump out hundreds of terabytes of imagery every day, far beyond what any analyst corps or rule‑based software can triage in real time. Deep‑learning pipelines such as ICEYE’s convolutional flood classifiers and OroraTech’s FireDetect™ transformer sift noise from signal in milliseconds, flagging a 25 m² hotspot before smoke is visible and estimating flood depth to actionable precision. Classical post‑processing would take days, rendering the UN’s early‑warning mandate impossible. Without AI, revolutionary hardware would decay into an unread archive.

Real‑time decisions

Governments can evacuate earlier and allocate crews with surgical precision. Google’s Flood Hub and related early warning initiatives in India have delivered flood alerts to millions of people, improving local preparedness and response.⁵

Insurance:
Insurers are increasingly using rapid-impact satellite maps, such as ICEYE flood data, to develop parametric insurance policies that can speed up claims settlements. MAPFRE RE’s agreement to license ICEYE flood data is intended to accelerate claims workflows and reduce settlement times. ⁶

Enterprise:
Timber firms and utilities are using near-real-time satellite analytics to improve environmental risk management and demonstrate ESG performance. These tools support efforts to track burn areas and optimise operations in regions such as Bavaria.

Society at large wins most. When Rio Grande do Sul flooded in 2025, SAR maps let officials funnel aid to vulnerable families within days, proof that fast data can translate directly into faster relief.⁷

Market and ethics

ICEYE fields the largest commercial SAR fleet (48 satellites today; 54 planned by year‑end). OroraTech is sprinting toward a dedicated, always‑on thermal network. Planet and Maxar keep optical dominance; Capella edges up on high‑res SAR. Yet speed alone is not a moat; the real edge lies in algorithms that convert raw pixels into trusted decisions in minutes.

Acceleration raises new questions. Models can under‑predict in data‑poor regions, and cross‑border imagery still triggers sovereignty debates. Transparent accuracy scores, bias audits and privacy controls must travel with every API call if the sector is to keep public trust.

The coordination layer

Insight becomes impact only when thousands of dispersed actors share the same picture fast enough to act. Emerging coordination platforms, including Stelia, ingest multi‑sensor feeds and surface a single operational view to responders and insurers alike. The result is a shift from reactive recovery to proactive resilience, delivered over secure, enterprise‑grade rails.

Digital‑twin forecasting

The next leap is already taking shape. Europe’s Destination Earth programme is fusing satellite streams, super‑computing and AI to build a live digital twin of the planet, able to “test‑fly” disaster scenarios before they happen.⁸ On the modelling front, DeepMind’s GraphCast forecasts global weather up to 10 days ahead with record accuracy, running in minutes on commodity chips.⁹ When these AI engines merge with near‑real‑time EO, decision‑makers won’t just see hazards sooner; they’ll simulate response plans, budget impacts and supply‑chain knock‑ons before the first raindrop or ember lands.

Build the warning network

The UN’s 2027 target for universal early warnings will not be met by satellites alone. It will take shared standards, interoperable APIs and boundary‑less collaboration. Platforms that harmonise diverse data sources into decisive, trustworthy intelligence are the missing layer.

Stelia is committed to partnering across that ecosystem, helping turn today’s patchwork of orbital eyes into tomorrow’s global nervous system. Because the world already has the eyes in the sky; what it needs now is the brain that turns sight into safety.

References

1. Natural Disasters in 2024 – Full‑Year Fact Sheet, Munich Re, January 2025. Climate change is showing its claws: The world is getting hotter, resulting in severe hurricanes, thunderstorms and floods | Munich Re

2. Early Warnings for All, United Nations. https://www.un.org/en/climatechange/early-warnings-for-all

3. Flood Rapid Impact Launch, ICEYE press release, July 2025. ICEYE unveils machine learning-powered Flood Rapid Impact Product to revolutionize response

4. OroraTech Wildfire Constellation, eoPortal, June 2025 OroraTech Wildfire Constellation - eoPortal

5. AI for Reliable Flood Forecasting, Google blog, March 2024. How we are using AI for reliable flood forecasting at a global scale

6. MAPFRE RE Licenses Global Flood Data, ICEYE press release, July 2025. MAPFRE RE signs agreement to license ICEYE’s global flood data

7. Supporting Rio Grande do Sul Government, ICEYE press release, July 2025. ICEYE supports Rio Grande do Sul government in mapping flooded areas and mobilizing humanitarian aid

8. Destination Earth: The Digital Twin Helping to Predict – and Prevent – Climate Change, IT Pro, July 2025. Destination Earth: The digital twin helping to predict – and prevent – climate change

9. GraphCast: AI Model for Faster and More Accurate Global Weather Forecasting, DeepMind blog, November 2023. GraphCast: AI model for faster and more accurate global weather forecasting

   Subscribe now

What if you could generate a synthetic version of every patient in a clinical trial — one that shows exactly what would have happened to them without treatment?

Unlearn.ai built exactly that system. Using conditional restricted Boltzmann machines, they create counterfactual patient trajectories that reduce trial sizes by 30–50% without losing statistical power.

The engineering challenge isn’t just building generative models. It’s doing so with medical data that has 15–40% missing values and irregular temporal sampling, while meeting regulatory requirements that would break most ML pipelines. The result is a production system that fundamentally changes how clinical evidence gets generated — but building it required solving problems that don’t exist in typical ML deployments.

The architecture: conditional generation at scale

Unlike standard ML models that predict P(outcome|features), this system learns P(outcomes | baseline_features, treatment_assignment, time_horizon) — essentially building a time-series generator conditioned on medical context. Think of it like dependency injection for probabilistic models: instead of hardcoding assumptions about patient populations, you inject specific patient contexts at runtime.

The breakthrough methodology, called PROCOVA (Prognostic Covariate Adjustment), transforms the entire statistical inference process. Instead of comparing treatment groups against population averages, each patient gets compared to their digital twin:

# Traditional approach
P(outcome | features)

# Unlearn.ai's conditional generation
P(outcomes | baseline_features, treatment_assignment, time_horizon)

# Per-patient comparison
Treatment_Effect_i = (Observed_Outcome_i - Predicted_Natural_History_i) / Predicted_Natural_History_i

This shifts from a between-group comparison problem to a within-patient prediction problem. The baseline heterogeneity that creates noise in traditional trials becomes useful signal for personalised modelling — turning what was previously statistical noise into valuable training data.

The conditional restricted Boltzmann machine architecture splits its internal representation across different timescales. Similar to how modern systems separate hot and cold storage paths, separate hidden unit populations capture short-term clinical fluctuations versus long-term disease progression — like having different model components specialised for high-frequency and low-frequency patterns in the same probabilistic framework.

Training requires sliding temporal windows across massive datasets: electronic health records spanning decades, clinical trial databases, real-world evidence from insurance claims. Each training example effectively answers: “Given this patient profile at this time point, here’s how their disease progressed over the following months.”

The cRBM training pipeline resembles large language model pretraining — massive datasets, GPU-accelerated optimisation, generation of synthetic outputs that preserve statistical structure from training data. The critical difference is conditioning and consequence: where language models autocomplete text, this system forecasts disease trajectories with direct clinical and regulatory implications.

Data engineering reality check

This conditional generation approach sounds elegant in theory. But like most ML systems, the mathematics are only as good as the training data — and medical data presents failure modes that would cripple models trained on cleaner domains like computer vision or natural language processing.

When a patient’s lab results are missing, it could be a clinical decision (test wasn’t ordered), operational failure (test was ordered but not performed), documentation gap (test performed but not recorded), or technical failure (test recorded but not transmitted). Each scenario carries different implications for model training, because treating systematic gaps as random missingness introduces bias that propagates through every generated trajectory.

Like handling eventual consistency in distributed databases, you can’t assume all medical data arrives on time or at all, so the system must gracefully handle partial information. The platform implements multi-layered data validation that operates continuously rather than as a preprocessing step:

This approach builds quality assessment directly into the generative process so that model uncertainty reflects actual data reliability rather than algorithmic confidence — ensuring that every synthetic patient trajectory is grounded in reliable medical evidence.

Production-scale engineering challenges

With reliable data pipelines established, the next challenge swiftly arises: generating thousands of synthetic patient trajectories in real-time while maintaining the statistical validity that regulators demand. This creates computational and memory constraints that don’t exist in typical ML deployments.

Distributed computation: The architecture distributes inference across GPU clusters, with each processing unit handling distinct patient cohorts. Shared model parameters ensure consistent probabilistic behaviour while enabling high-throughput trajectory generation. The parallelisation strategy accounts for the fact that patient trajectories can be generated independently once the model is trained.

Memory-efficient state management: Long-horizon simulations for conditions like cardiovascular disease can span two-year periods with dozens of clinical variables per patient. The system minimises memory overhead through trajectory compression and selective caching strategies that store only clinically significant state transitions.

Adaptive trajectory updates: As real trial data streams in, each digital twin incorporates new information to refine its predictions. A patient’s unexpected lab result or adverse event updates their model state, adjusting the synthetic trajectory moving forward. This creates a feedback loop between trial execution and generative interface — unlike typical batch prediction systems that operate on static datasets.

The regulatory requirements create constraints that most ML systems never face:

Explainability requirements: While cRBMs excel at capturing complex statistical dependencies, they don’t naturally produce the interpretable outputs that regulators require. The platform implements gradient-based attribution methods that trace outcome predictions back to specific baseline variables and prior clinical events, building transparency into a fundamentally generative system.

Technical validation at scale

All this infrastructure investment raises a critical question: how do you prove that synthetic patients are statistically equivalent to real ones? The validation approach operates under a simple principle: model sophistication cannot compensate for data unreliability, so continuous validation monitors multiple dimensions simultaneously.

This resembles chaos engineering for ML models — continuously testing edge cases and failure modes rather than assuming training performance generalises to production scenarios:

Primary validation:

• Data drift detection: Real-time comparison of incoming patient data against training distribution parameters. When new patients exhibit characteristics outside the model’s training envelope, the system flags potential extrapolation errors and adjusts confidence intervals accordingly.

• Cross-population validation: The platform maintains separate validation datasets for different demographic groups, disease stages, and treatment protocols. Before generating synthetic trajectories for any patient cohort, the system verifies that adequate training data exists for that specific population segment.

Continuous monitoring: For ongoing trials, the system continuously compares its synthetic predictions against actual patient outcomes as they become available. Significant deviations trigger automated investigation into whether the discrepancy reflects model limitations, data quality issues, or genuine clinical surprises that require model updates.

Statistical calibration: High statistical power only provides value if prediction confidence is properly calibrated across diverse patient demographics and disease stages. The platform continuously tests and adjusts its uncertainty estimates, ensuring confidence intervals reflect actual prediction accuracy rather than model confidence.

Each of these validation capabilities serves the larger goal of reducing trial sizes while maintaining statistical rigour — proving that synthetic patients can reliably substitute for real ones in clinical research.

Quantified engineering impact

This comprehensive validation framework enables Unlearn.ai to make a bold claim: their technical improvements translate into measurable changes in trial efficiency.

PHASE II ONCOLOGY TRIALS
Typical: 200-400 patients → With Digital Twins: 120-280 patients
↓ 30-40% reduction

PHASE III CARDIOVASCULAR TRIALS  
Typical: 2000-4000 patients → With Digital Twins: 1400-2800 patients
↓ 30-50% reduction

TIMELINE IMPACT: 25-40% faster completion
STATISTICAL POWER: Maintained at equivalent levels

Broader engineering implications

This system demonstrates how generative modelling built on conditional probability structures enables not just inference, but simulation. It moves beyond prediction into counterfactual reasoning at population scale — a pattern that opens applications far broader than medical AI.

The technical achievement lies not in abstract algorithmic complexity, but in solving concrete real-world constraints through generative computation. The mathematics transforms a fundamental limitation in how medical evidence is generated, while the engineering makes it operationally viable at the scale required for regulatory acceptance.

For developers working on other domains requiring counterfactual reasoning — recommendation systems that need to model alternative user journeys, financial models that simulate different market conditions, or any system where you need to answer “what would have happened if” — this architecture provides a production-tested pattern for conditioning generative models on complex, multi-dimensional contexts while maintaining statistical rigour at scale.

Subscribe now

By 2030, more than one in six people globally will be over 60. Healthcare costs are climbing 5% annually, outpacing GDP growth in developed nations. Meanwhile, non-communicable diseases account for 74% of global deaths, yet early detection and prevention remain limited by a fundamental constraint: our inability to understand disease origins at the molecular level quickly enough to matter.

The promise of genomics has existed for decades. Sequence a person's DNA, understand their biological blueprint, intervene before disease takes hold. The science works. What hasn't worked is the operationalisation. Traditional genomic analysis takes weeks, costs thousands, and requires specialised facilities accessible to a fraction of the global population.

This gap between scientific possibility and practical accessibility defines one of the most pressing challenges in healthcare today. It's also becoming a defining test case for AI's role in addressing planetary-scale human needs.

The speed revolution

A Canadian biotechnology company called BioAro recently achieved something that reframes this entire discussion. Their PanOmiQ platform completed whole genome sequencing analysis in under two hours, generating variant call format files in under five minutes. To put this in perspective: the first human genome took 13 years to sequence. Current industry standard is about four weeks. BioAro compressed this to seven hours with 100% accuracy verified by the College of American Pathologists.

This isn't just faster computing. Speed at this scale transforms what genomic medicine can be. Emergency room diagnoses. Rural clinic interventions. Real-time treatment decisions in developing nations. When analysis time drops from weeks to hours, genomic medicine stops being a specialised service and starts becoming standard care.

The technical achievement rests on AI systems that can process multi-omics data in real time, integrating genomics, proteomics, metabolomics, and microbiome analysis into unified insights. More significantly, these systems are designed for global deployment from the start, generating clinical reports in six languages and operating across diverse healthcare infrastructures.

This represents a pattern emerging across AI applications: the organisations succeeding at scale aren't just building better algorithms. They're building systems architected for planetary-scale distribution from day one.

Democratisation through AI

The implications extend far beyond genomic analysis speed. BioAro's platform eliminates the high costs and complexity that have historically made whole genome sequencing accessible only to major medical centres. When AI can automate what previously required teams of specialists, advanced genomics becomes deployable in hospitals, research institutions, and underserved communities worldwide.

This democratisation effect illustrates something crucial about AI's potential impact on humanity. The technology's greatest contribution may not be its raw capabilities, but its ability to make sophisticated human knowledge accessible at scale. Complex expertise that once required years of training and expensive infrastructure can now be embedded in systems that operate reliably across diverse environments and user contexts.

The multilingual reporting capability provides another glimpse of what global AI deployment requires. Building systems that work across cultural and linguistic boundaries demands more than translation. It requires understanding how medical knowledge is communicated, interpreted, and acted upon in different healthcare contexts. Success at planetary scale means accounting for this diversity from the architectural level, not adding it as an afterthought.

Organisations attempting to operationalise AI globally are discovering that the technical challenge of making algorithms work is often simpler than the human challenge of making them work everywhere, for everyone.

From code to cure

BioAro has also introduced something they call an "omics-to-therapeutics" ecosystem. Their AI doesn't just analyse genetic data; it identifies novel therapeutic targets and uses both structure-based and ligand-based modelling to predict how potential drugs might behave in the human body. The system generates chemical candidates from biological insights at what they describe as "machine speed."

This integration reveals another dimension of AI's potential impact. Rather than replacing human expertise in drug discovery, the platform amplifies it by handling the computational heavy lifting that currently creates bottlenecks. Researchers can focus on the creative and strategic aspects of therapeutic development while AI manages the systematic analysis of massive biological datasets.

The economic implications are substantial. Traditional drug discovery takes decades and costs billions, with high failure rates. When AI can compress the timeline from biological insight to chemical candidate, it changes the fundamental economics of developing treatments for rare diseases, personalised therapies, and conditions that affect populations in developing nations.

This points to a broader principle: AI's most transformative applications may emerge when it enables entirely new approaches to existing problems, rather than simply optimising current processes.

Redefining human potential

What becomes possible when genetic analysis shifts from a specialised service to a standard capability available anywhere healthcare is practised? The answer goes beyond faster diagnoses or more targeted treatments.

Predictive medicine becomes feasible at population scale. Instead of waiting for symptoms to appear, healthcare systems can identify genetic predispositions and intervene before diseases develop. This shift from reactive to preventive care could fundamentally alter how we think about health and ageing.

The microbiome analysis capabilities built into BioAro's platform illustrate another frontier. By understanding the interplay between genetics and the microbial ecosystems in our bodies, AI enables interventions tailored not just to our genes, but to our complete biological context. Personalised nutrition, targeted immunotherapy, and chronic disease management all become more precise when guided by comprehensive biological data.

These advances point towards a future where medicine becomes truly individualised. Rather than treating populations with standardised protocols, healthcare providers can design interventions specific to each person's unique biological profile.

The challenge is ensuring these capabilities benefit everyone, not just those with access to advanced medical systems.

The global deployment challenge

Achieving this vision requires solving what may be AI's defining challenge: moving from promising laboratory results to reliable performance across diverse real-world environments. This is where many AI initiatives stumble. Systems that work brilliantly in controlled settings often struggle when deployed globally.

BioAro's approach offers lessons for organisations grappling with this transition. Their platform is designed for multiple deployment models, supporting both cloud-based and on-premise implementations. This flexibility allows healthcare systems with different technical infrastructures and regulatory requirements to adopt the technology in ways that work for their specific contexts.

The quantum computing integration they're developing represents another approach to scale-out challenges. By leveraging quantum processing for genomic analysis, they're building systems that can handle exponentially larger datasets while maintaining real-time performance. Beyond commodity computational power, creating the coordination layer needed to serve billions of people simultaneously is the grander challenge.

Organisations attempting to deploy AI at planetary scale consistently encounter this same fundamental challenge: how to build systems that work reliably across the full spectrum of human contexts, technical environments, and cultural frameworks.

The time is now

The transformation BioAro is driving in genomics illustrates what becomes possible when complex systems are designed with both scientific rigour and human impact in mind. At Stelia, we see this as part of a broader movement - one that recognises AI’s future isn’t just technical, but societal, and that true progress lies in harmonising complex systems with collective wellbeing. The organisations succeeding in this transition share common characteristics: they prioritise deployed systems over promised capabilities, integrate diverse perspectives into their development processes, and focus on measurable outcomes rather than aspirational claims.

BioAro’s work reflects a deeper shift already underway - one where the value of AI is measured not just by what it can optimise, but by the real-world outcomes it enables for human lives. This is no longer just a question of technical acceleration, but of ethical integration.

The future belongs to those who can architect for both.

Subscribe now

Schmidt, super‑intelligence, and sovereignty

Last week former Google chief executive Eric Schmidt spent seventy‑five minutes on the Moonshots podcast with Peter Diamandis and Dave Blundin. The headline moment came when he declared that digital super‑intelligence is “within ten years.” The other take‑away: those future models, he said, will live in giga‑scale data centres guarded like nuclear stockpiles.

The remark has landed because policy circles from Washington to Brussels are already debating “sovereign AI”. If national power is set to hinge on who controls the biggest models, Schmidt’s comments sound like a strategic field manual. They also, as this article shows, rest on a shaky reading of how machine‑learning systems actually spread.

Schmidt’s vision in a nutshell

“If the structure of the world in five to ten years is ten models – five in the United States, three in China, two elsewhere – those data centres will be all nationalised in some way. In China they will be owned by the government. The stakes are too high.” (31 min 57 s – 32 min 28 s)

Schmidt’s argument has four main pillars:

1. Ten‑year countdown
   “When do you see what you define as digital super‑intelligence? – Within ten years.” (1 h 22 m 09 s)

2. Energy as the bottleneck
   “The natural limit is electricity, not chips… The United States will need ninety‑two gigawatts more power.” (2 m 13 s – 3 m 46 s)

3. Nation‑scale data centres as strategic assets
   A training run for xAI’s Grok used a “ten‑billion‑dollar super‑computer in one building.” (34 m 05 s – 34 m 34 s)

4. Deterrence through “mutual AI malfunction”
   A state‑level balance of power where each side can crash the other’s mega‑centre if lines are crossed (25 m 56 s – 27 m 12 s).

In this frame sovereignty equals location: keep the decisive compute inside your borders and guard it with armed security.

Two AI races, one sovereignty problem

Eric Schmidt actually sketches two separate competitions.

Race 1: the fortress sprint
“These data centres will be nationalised in some way… the stakes are too high.” (31 min 57 s)
In other words, only a handful of states or mega‑firms can afford a gigawatt hall and a ten billion‑dollar training run.

Race 2: the diffusion scramble
Minutes later he warns that, once the weights are distilled, “the final brain can run on four or eight GPUs – a box about this size,” making powerful models “a hundred, a thousand, a million” in number (33 min 19 s; 41 min 50 s).

During the episode Dave Blundin recalls a visit to OpenAI where researcher Noam Brown (not on the podcast) predicted that models will soon write their own step‑by‑step workplans, or “scaffolding,” making multi‑hour tasks autonomous by 2025. If true, diffusion in Race 2 speeds up because the planning layer travels with the weights, not with the cloud provider.

Why this matters for ‘sovereign AI’

• Fortress sovereignty may work during Race 1 while the capability lives in guarded sites.

• Race 2 dissolves that physical anchor. Portable checkpoints flow across borders faster than regulators can stamp passports.

• Schmidt offers a deterrence doctrine for Race 1 but admits the diffusion phase is “a set of unknown questions.”

• Location‑based sovereignty will have little time to prove its worth before portable, self‑planning checkpoints flood every jurisdiction.

Stelia calls this the sovereignty paradox: legal control framed around location collapses at the very moment the technology scales to everyday hardware.

Three gaps in the Schmidt approach

1. Compute does not equal control

Schmidt’s fortress protects the factory, not the product. Moments after describing armed guards he concedes that the same model, once distilled, can run on a desktop server:

“The final brain can be ported and run on four or eight GPUs – a box about this size.” (41 m 50 s – 42 m 25 s)

When weights leave the building they lose all jurisdictional tags. The link between physical residency and real control vanishes.

2. Deterrence needs a target – what if there is none?

Nuclear doctrine worked because silos were fixed on maps. Schmidt’s “mutual AI malfunction” assumes similar visibility. Yet portable checkpoints allow powerful models to live on laptops, research clusters, open‑source mirrors and even criminal botnets. A strategy that threatens single buildings cannot restrain actors who own none.

3. Whose sovereignty counts?

The Schmidt narrative is Washington versus Beijing. Missing are Indigenous communities whose language archives feed multilingual models, European citizens protected by GDPR, or African start‑ups seeking local agency. National control over hardware does not answer their claims to participation and benefit.

A governance path that could work

Stelia’s engineers operate multi‑jurisdiction deployments every day. Three design principles align with technical reality:

Residency first: Keep raw data and first‑training runs inside legally recognised regions using geo‑fenced compute and region‑tied encryption keys.

Transparency of weights: Publish weight hashes and lineage metadata so any checkpoint can be traced, wherever it runs. Add inference watermarks for audit.

Participation hooks: Build consent, veto rights and benefit‑sharing into dataset pipelines so communities influence use before models go live.

These tools do not pretend to project domestic law across borders. Instead they create verifiable constraints that travel with the artefact itself – a sharper fit for how machine‑learning actually works.

Why this matters now

With Race 1 already under way and Race 2 close behind, the window for workable governance is short.

“We are saying it is 1938… we need to start the conversation now, well before the Chernobyl events.” (28 m 20 s – 28 m 41 s)

If policymakers chase the mirage of fortress sovereignty, genuine levers of control will ossify behind optics and compliance theatre. Residency, transparency and participation offer a route that communities can verify and industry can implement.

Stelia will publish a technical white paper on data sovereignty later this month.

Subscribe now

Source: Moonshots with Eric Schmidt (YouTube, June 2025): Ex-Google CEO: What Artificial Superintelligence Will Actually Look Like w/ Eric Schmidt & Dave B

AdTech has reached a critical disconnect. As AI models become more sophisticated at predicting user behaviour and optimising campaigns, the data required to train these systems becomes increasingly restricted by privacy regulations. Centralised cloud AI collides with per‑country data rules, creating a fundamental mismatch¹.

The compliance trap

This creates what industry veterans are calling the "compliance trap" – where the most effective AI strategies become legally impossible to implement using legacy cloud platforms.

Recent policy developments underscore how profoundly policymakers misunderstand this challenge. Europe’s €200 billion ‘AI gigafactory’ keeps training inside its borders. Yet a finished model is just numbers; once copied, it operates anywhere. This reflects what researchers term the "sovereignty paradox": the more sophisticated AI systems become, the less meaningful traditional geo‑territorial control mechanisms prove to be².

For AdTech, this paradox reveals a fundamental flaw in compliance strategies based on geographical data control. A trained targeting model is simply a matrix of numbers. Once training completes, those weights can be perfectly copied and deployed on smartphones, connected TVs, and programmatic platforms worldwide, entirely beyond the jurisdiction where training occurred.

How federated learning actually works

To understand why this matters for AdTech, consider how federated learning solves the fundamental data‑access problem. Traditional AI training requires collecting all user data in one central location, training a model on the combined dataset, then deploying globally. This approach risks breaching local privacy legislation, creates security risks, and requires expensive data movement across jurisdictions.

Federated learning flips this model entirely. Instead of moving data to where computation happens, it sends models to where data lives. Here's the process for a global AdTech platform wanting to improve click prediction across EU and US markets:

Clone – A central server creates an initial click‑prediction model and distributes identical copies to EU data centres (Frankfurt) and US data centres (Virginia).

Train locally – The EU facility trains the model on German, French, and Italian user data locally, while the US model trains on American user data locally. Raw user data never leaves its regional boundaries.

Send weights – After training, each data centre sends back only model parameters – mathematical weights and biases that represent learned patterns.

Merge – The central coordinator combines EU and US updates using algorithms such as Federated Averaging, creating an improved global model without moving personal data across borders.

Re‑deploy – The updated model returns to both regions and the cycle repeats for continuous improvement.

The compliance reality

This architecture directly addresses core compliance challenges. The EU centre processes millions of EU users' clicks and sends only weights to the coordinator; the US centre does the same with American data. What flows are mathematical insights, not personal information³.

The European Data Protection Board’s Opinion 28/2024 clarifies that parameter exchange aligns with GDPR’s data‑minimisation and purpose‑limitation principles⁴. In 2023, Ireland’s Data Protection Commission fined Meta €405 million for GDPR violations tied to behavioural advertising⁵. Federated learning offers a potential privacy‑enhancing alternative, keeping raw data local while enabling global optimisation.

Legacy platform limitations

General‑purpose clouds support federated learning only as an add‑on, not a design principle. AWS, Azure, and Google Cloud each provide federated toolkits, yet they rely on centralised foundations that re‑introduce compliance vulnerabilities and add DevOps overhead⁶.

The mathematical reality makes geographical training control meaningless. A personalised ad model trained on European infrastructure is identical to any copy of its numerical weights. Once deployed globally, these models operate beyond territorial governance, turning “Sovereign AI” into compliance theatre. Standard cloud tools also struggle with multi‑party coordination; authentication and synchronisation often fail when a TV OEM, streaming service, and programmatic platform attempt joint training.

Technical performance reality

Federated learning delivers up to 30 % higher accuracy because local data diversity stays intact. Traditional centralised training homogenises insights, whereas federated approaches preserve regional signal and enable nuanced targeting, crucial in contexts like connected‑TV viewing, which varies by culture, device, and content genre.

The coordination challenge

Successful adoption requires orchestrating training across many independent systems. Identity management, version control, and rollback mechanisms must be standardised to avoid conflict and ensure auditability, needs that exceed the scope of most legacy ML tooling.

Strategic implementation necessity

Regulations are tightening while third‑party data dwindles; the gap between distributed and centralised AI will widen. Valued at $151 million in 2024, the federated‑learning market is projected to reach $507 million by 2033 (CAGR 13.6 %)⁷. By embedding governance directly into system architecture, federated learning addresses real regulatory risk rather than relying on geographic workarounds.

Partnership strategy is crucial: consortiums of streaming services, publishers, and measurement vendors can pool insights while sharing costs. The real question is which organisations will secure first‑mover advantages before legacy approaches become untenable.

Organisations that build federated capabilities now gain a sustainable edge that competitors will struggle to replicate as regulations harden and competitive pressures intensify. The compliance trap is real, but so is the solution for those prepared to implement governance through architecture rather than geography.

Next steps for decision‑makers

• Evaluate existing data flows against current and pending regulations.

• Run a limited federated pilot (e.g., one EU market, one US market) to measure performance uplift and confirm compliance reporting.

• Establish a cross‑functional team combining legal, data science, and platform engineering to formalise governance processes.

For organisations ready to explore this approach in detail, a short technical assessment can clarify requirements, expected ROI, and integration timelines.

Subscribe now

References

¹ Opinion 28/2024: Certain Data Protection Aspects of the Use of AI Systems in the EU, European Data Protection Board (EDPB), June 2024.
https://www.edpb.europa.eu/our-work-tools/our-documents/opinion-board-art-64/opinion-282024-certain-data-protection-aspects_en

² EDPB Opinion 28/2024 – Territorial Limits of AI Model Governance, European Data Protection Board (EDPB), December 2024.
https://www.edpb.europa.eu/system/files/2024-12/edpb_opinion_202428_ai-models_en.pdf

³ Opinion 28/2024: GDPR Compliance and Parameter Exchange in AI Training, European Data Protection Board (EDPB), June 2024.
https://www.edpb.europa.eu/our-work-tools/our-documents/opinion-board-art-64/opinion-282024-certain-data-protection-aspects_en

⁴ EDPB Guidance: How AI Parameter Sharing Aligns with GDPR Principles, European Data Protection Board (EDPB), 2024.
https://www.edpb.europa.eu/news/news/2024/edpb-opinion-ai-models-gdpr-principles-support-responsible-ai_en

⁵ Irish DPC Fines Meta €405M for Instagram GDPR Breaches, Burges Salmon, September 2022.
https://www.burges-salmon.com/articles/102hxck/irish-data-protection-commissioner-fines-meta-405m-for-violation-of-childrens-p/
http://www.dataprotection.ie/en/news-media/press-releases/data-protection-commission-announces-decision-instagram-inquiry

⁶ Federated Learning Systems: Challenges with Multi-Party Coordination, arXiv preprint, February 2025.
https://arxiv.org/html/2502.05273v1

⁷ Federated Learning Market Report 2024–2033, IMARC Group, 2024.
https://www.imarcgroup.com/federated-learning-market

The illusion of progress

For nearly a decade, artificial intelligence has been touted as the future of media and entertainment. Netflix has invested heavily in AI-driven personalisation and operational systems, continuously testing new models for engagement and optimisation. Disney continues to test automation tools. Warner Bros. runs ongoing proof-of-concept projects. Yet for all the headlines, panels, and prototypes, meaningful integration has remained frustratingly elusive. Endless pilots and proof of concept projects have filled the last few years each full of promise, most ending in silence. Now, as I work alongside companies facing intensified pressure for profitability, audience retention, real-time personalisation, and operational agility, it is clear that isolated AI experiments are no longer enough. The pilot era is over. What comes next is intelligent orchestration, an approach fundamentally different, inherently more strategic, and critical for survival in a rapidly evolving marketplace.

The trap of pilot purgatory

A recent industry survey conducted in June 2025 painted a sobering picture. Despite rising AI budgets, over 60 percent of media executives are still stuck in pilots, spending more time testing than transforming and only 4 percent operationalised AI at scale¹. Even more telling, nearly half of industry leaders cited unclear use cases and ROI uncertainty as the largest barriers to effective adoption. I’ve seen this first-hand. At Cannes this year, I heard the same refrain across panels and beachside meetings: ‘We’re testing, but not scaling.’ But testing without commitment creates stagnation. This is what I call pilot purgatory, an endless loop of testing, proving, and pausing. It’s safe. It’s structured. But it doesn’t build anything. And in a landscape defined by rising costs, real-time engagement, and global pressure, staying in purgatory is not a neutral choice. It’s a risk.

Why fragmented AI isn’t enough anymore

Strategic pressure is mounting. In my work with content and technology teams, I hear it constantly. Streaming platforms are battling churn and the growing expectation for hyper-personalisation. Game publishers are navigating the complexity of live-service operations, compliance, and localisation. Production teams are buried in rights tracking and versioning. And agencies? They’re sprinting to deliver market-specific creative faster than ever. Fragmented workflows and isolated automation just can’t keep up. While platforms like TikTok and Meta are rolling out AI-driven DIY creative tools aimed at small businesses, the companies I work with don’t need more tools. They need orchestration with intelligence. They’re not looking to just generate more content. They want systems that adapt, decide, and deploy at scale.

The new runway for competitive advantage

What makes this moment different is not just that AI is better. It’s that the infrastructure to support true orchestration is finally emerging. Until recently, there were no systems built to connect content, data, compliance, and logic in real time across the full lifecycle of media creation and distribution. But that’s changing. AI native platforms are now being designed for orchestration. Built from the ground up to enable intelligence that is secure, governable, and affordable.

Beyond automation: the rise of agentic orchestration

The media and entertainment industry’s next evolutionary step is obvious: a strategic transition from isolated automation to systemic orchestration. Unlike traditional automation, which assists, agentic systems act. They decide, execute, adapt and respond across multiple workflows simultaneously without waiting for instruction. They are fit for a world that moves in real time.

When designed well, agentic orchestration becomes the intelligent backbone of a company’s operations. Not an add-on, not a tool, but the infrastructure through which decisions flow. Orchestration enables streaming platforms to dynamically tailor promotional content instantly across dozens of international markets, cutting localisation from weeks to hours. In gaming, real-time live-service updates, content translation, and compliance happen automatically, boosting player engagement while reducing friction. Film and audio production companies utilising orchestration gain real-time management of compliance, rights tracking, and asset metadata, significantly accelerating production timelines and reducing operational risk. Creative agencies implementing agentic orchestration rapidly iterate campaign assets, optimising creative output and minimising repetitive manual labour, freeing teams to concentrate on strategic innovation.

This is how media organisations move from friction to flow. From linear content operations to adaptive ecosystems that scale with demand.

The real shift is in leadership

But orchestration requires more than infrastructure. It demands a new kind of leadership. I often tell executives this: the real leap is not building smarter systems but becoming a different kind of leader. One who doesn’t manage tasks but designs systems that manage themselves. The most powerful leaders aren’t asking, ‘What should we automate?’ They’re asking, ‘What can I offload so I can focus on what actually moves us forward?’ What I can observe is that companies with high AI maturity get even 3× higher ROI than those just testing the waters.

Today’s CROs are no longer just chasing CPMs. They’re re-engineering revenue operations through agentic systems that adapt in real-time to audience, regulation, and creative context. The most effective COOs aren’t managing workflows they’re architecting feedback loops. And CMOs, if they’re smart, are investing in infrastructure that doesn’t just push creative, but shapes it through intelligence.

Governance, compliance, and performance monitoring are no longer separate layers. They must be embedded directly within orchestration frameworks. In every project I advise, we insist on real-time auditing, automated metadata, and rights tracking — because trust can’t be an afterthought.

Real-time feedback loops are integral, continuously feeding operational performance data, market signals, and audience insights into orchestration systems. These loops ensure constant optimisation and refinement of workflows, enabling rapid adaptation to changing market dynamics and customer expectations.

One overlooked shift is the need to shape brand presence not just for consumers, but for the AI agents now parsing the internet. Training data matters. Influencing LLMs may soon require content designed not to be read, but to be remembered by algorithms. Orchestration is how that happens.

Adopting orchestration strategically involves incremental implementation and rapid scaling. Initial pilot orchestration projects should target high-impact operational areas, clearly demonstrating value to build internal momentum and confidence. Modular integration of orchestration technology into existing infrastructures allows rapid deployment without large-scale disruptions, ensuring immediate and measurable operational improvements.

Strategic patience and iterative agility are equally essential. Leaders must communicate incremental successes transparently, remain patient with evolving results, and continually refine orchestration frameworks to align with shifting industry landscapes and organisational priorities.

The rise of agentic AI is also reshaping how we think about talent. As digital labour becomes embedded, leaders must invest in AI literacy, workflow design, and soft skills to make the most of these systems. Technology alone doesn’t transform companies, people trained to wield it do. This new orchestration should liberate talent, empower teams to shift from routine operations to strategic creativity, innovation, and storytelling, thus amplifying human creativity rather than replacing it.

Where to begin

The temptation is always to start small. A workflow. A tagging system. A campaign tool. I would say: don’t start with isolated tools. Instead, pick one high-friction area where multiple systems collide and you can deliver exponential returns. Orchestration only creates real value when it addresses connected complexity. For media and entertainment, the most powerful starting points are content localisation, live-service updates, multi-region rollouts, and rights-bound metadata.

As generative agents begin to shape not just content creation but content discovery, we’re entering a phase where orchestration needs to consider not only human workflows, but AI-to-AI ecosystems. The next recommendation engine might not be a human but an LLM choosing one brand over another.

These orchestration pipelines must be modular, governed, and integrated with real-time data sources. They must evolve as the audience, platforms, and content evolve. They must be designed not as tools, but as systems.

Not just faster: Smarter. Freer. Future-Ready.

The companies that embrace this shift won’t just move faster. They’ll operate differently. With less friction. With more clarity. And with systems that scale intelligence as easily as they scale content. The next decade in media and entertainment won’t be defined by who tests AI the longest. It will be shaped by who orchestrates it best. The pilot culture is over. What we build next is what will last.

by Ula Nairne, VP Media & Entertainment, Stelia

References

¹ AI Adoption in the Media Industry: Where It’s Working and Where It’s Stuck, Amagi & Dan Rayburn, June 2025. https://www.amagi.com/resources/insights-ai-impact-media-industry

Subscribe now

Let’s not sugarcoat it; the state of AI governance is a mess. Worse yet, it’s showing no signs of getting better anytime soon.

Just this month (July 2025), Washington saw a dramatic clash over the future of AI regulation. The Trump-aligned proposal to block states from enacting their own AI laws (a move washed in to the “Big, Beautiful Bill”) was flatly rejected in the Senate. The message from lawmakers — States should be free to protect their residents, even if it means a messy web of conflicting rules.

Meanwhile, over in Brussels, the European Union is forging ahead with its own AI Act, brushing aside rumours of delays to enactment and signalling its intent to lead on AI governance.

This transatlantic divergence should worry anyone paying attention. AI is a borderless, general-purpose technology, yet we’re attempting to regulate it with tools and mindsets built for a pre-digital world, fractured by political agendas and local priorities.

Lawmakers writing these AI regulations lack the technical understanding needed to grasp the technology they’re governing, ignoring highly experienced legal professionals and academics with years of experience in the sector. This raises the question: are these laws designed to protect people or simply to create frameworks governments can navigate? Without expert insight, regulation risks becoming symbolic, not practical and ultimately ineffective. This isn’t protection, this is chaos.

A patchwork that is already tearing

The current regulatory approach is a tangled mess; a chaotic patchwork of state-level and national laws, each with its own definitions, demands and deadlines. The problem isn’t just complexity, it’s incoherence.

AI doesn’t neatly respect industry boundaries, let alone state or national ones. A single model can be deployed globally, retooled across sectors and evolve faster than any regulator can track. Yet instead of crafting a unified framework, lawmakers are scrambling to patch together fragmented responses while the ground shifts beneath them.

In the US, proponents of the failed federal AI moratorium (including some prominent tech figures) warned that a labyrinth of 50 state regulations could cripple innovation and hand China a competitive edge. OpenAI’s own Sam Altman bluntly told senators that managing compliance across so many jurisdictions would be practically impossible.

But their warnings weren’t enough.

The AI freeze was stripped from the bill leaving America exactly where it was to begin with: fragmented, uncertain and without a national game plan.

The big AI bet

Across the pond the EU is betting big on regulation. The AI Act, modelled on the bloc’s approach to the GDPR, aims to impose a sweeping risk-based framework on AI systems used within its borders. On paper it sounds like progress with transparency for low-risk systems, tight controls for high-stakes use cases like hiring or policing and new rules for the looming threat of general-purpose AI (GPAI).

But the tech is evolving faster than the rules. A single AI model might help you write an email today and guide a cancer diagnosis tomorrow. Trying to classify these systems by risk level is like nailing jelly to an ever moving wall. And the Act’s GPAI provisions? Vague, reactive and already chasing a moving target with consultations still being undertaken on what GPAI even is (and we’ve not even got to artificial general intelligence yet…).

To their credit, EU lawmakers have acknowledged the problem and tried to adjust late in the game. But even as they push ahead, the uncomfortable truth remains: Europe’s AI rules may be outdated before the ink is dry.

One world, many rules

At the heart of the problem is a basic contradiction; AI is global, regulation isn’t.

A generative model might be trained in California, fine-tuned in London, hosted in Singapore, and used by someone in Nairobi. So whose laws apply? Whose values shape the outcomes? Who is accountable when things go wrong? And how does this even become ‘Sovereign’.

Right now there are no clear answers. The US and EU are each pursuing their own philosophies — decentralised experimentation versus centralised oversight. Other countries like the UK, are forging different paths altogether. And while there are talks at the G7 and the UN, global coordination remains little more than a diplomatic soundbite.

What we’re left with is a high-stakes experiment playing out in real time. The world is testing governance models in parallel, without consensus, without interoperability and without a safety net.

Time to wake up

Here’s the hard truth; local rules are no match for global systems.

AI is rapidly becoming foundational to how we live, work, make decisions and frankly exist. If we don’t find a way to align on basic principles of safety and accountability, we risk letting this technology outpace our ability to control it.

Yes, fragmented governance offers room for experimentation. Yes, different regions have different values. But let’s not pretend that a patchwork of uncoordinated rules is a long-term solution. At best, it’s a stopgap. At worst, it’s a recipe for regulatory chaos and ethical disaster.

The world doesn’t need one-size-fits-all regulation, but it does need a shared vision. Without it, we’re not governing AI, we’re just reacting to it. And sooner or later, the consequences of that passivity will catch up with us.

The question is no longer whether we need global coordination; it’s whether we’ll get there before it’s too late.

by Imogen Armstrong, Chief Legal Officer, Stelia

We asked our VP Engineering what’s the latest thing that’s made you stop and think: “wow that’s clever”. Here’s what he had to say:

“I tend to be pretty selective about what catches my attention.” David Hughes, Stelia VP Engineering, “Everyone these days is obsessed with ‘We don’t get enough tokens out of this GPU, let’s just buy a bigger, more powerful GPU.’ That’s not the answer. There’s an interesting bit of software from a MIT HAN Lab project at the minute called QServe that flips the whole problem on its head [1].”

Challenging Scale-Up Thinking

That frustration with brute-force scaling is exactly why QServe, the project Dave called out from MIT, stood out to him. QServe is an inferencing library developed by researchers at MIT, specifically focusing on reducing bottlenecks, and increasing throughput in LLM serving — a clear rejection of the “just throw power at it” mentality.

In real-world terms, quantisation can meaningfully accelerate LLM inference.

Quantisation as a Runtime Strategy

Traditionally, deep learning relies on precision formats like FP16, FP32, FP64, and INT8. As research has advanced, there has been a clear push beyond INT8, exploring even lower precision such as INT4. However, INT4 quantisation techniques have mostly been effective in low-batch, edge-style LLM inference, while struggling to deliver gains for large-batch cloud-based serving. That shortfall is primarily due to runtime penalties introduced during dequantisation of either weights or partial sums on GPUs.

To address this, QServe implements a W4A8KV4 quantisation algorithm (meaning 4-bit weight, 8-bit activation, 4-bit KV cache), following a 4–8–4 pattern known as QoQ. This approach achieves measurable speedup by recognising a key insight: the efficiency of LLM serving is often critically bottlenecked by operations on low-throughput CUDA cores. QoQ tackles this with progressive quantisation, lowering dequantisation overhead and pairs it with SmoothAttention to mitigate the accuracy losses typical of 4-bit quantisation. Additional strategies include compute-aware weight reordering, register-level parallelism to reduce quantisation latency, and making fused attention memory-bound to fully harness the performance gains of KV4 quantisation.

The Performance Picture

Benchmarks from the QServe team [1] show throughput gains of 1.2–1.4x on LLaMA-3–8B and up to 3.5x on Qwen-1.5–72B, with L40S GPUs even outperforming A100-class setups in several scenarios. These results tie back to QServe’s efforts to minimise dequantisation bottlenecks on GPU cores, which are a well-known source of INT4 inefficiency. Reported token costs were up to three times lower, thanks to reduced kernel overhead and higher hardware utilisation. That said, the paper notes remaining challenges around runtime stability under high concurrency and large context windows. It’s promising but not guaranteed to hold under every production scenario.

The Implementation Trade-Offs

Beyond consistency, this kind of adaptive quantisation also isn’t trivial to roll out. The entire inference stack — from tokenizers to kernel-level scheduling — needs to tolerate non-uniform precision at runtime. That demands deeper model introspection, more robust calibration workflows and a willingness to invest in kernel-level engineering, especially for workloads beyond the benchmarked standard models. It’s not a drop-in fix, but for anyone willing to trade a bit of complexity for massive efficiency gains, QServe offers a solid blueprint.

What History Teaches Us

Dave noted the similarities between QServe’s solution and earlier industry breakthroughs:

“In a throwback to ye olden days of the storage world, we were presented with (seemingly) hard limits of, for example, 1TB of physical storage — and yet we found ways to work around this in software. Enter: on-the-fly compression which, as algorithms developed further, became more and more efficient and you had different compression algorithms for different use cases, e.g. LZ4 vs GZIP and even continued to innovate in later years with ZSTD becoming even more efficient and delivering the sweet spots of both the aforementioned.

We will undoubtedly reach the same point in AI workloads, where we are presented with (again, seemingly) hard limits of the physical infrastructure involved and we will have a need to get clever. QServe is a glimpse of what getting clever might actually look like.”

Ultimately, QServe is a wake-up call about breaking habits. While the industry has been conditioned to worship bigger GPUs on reference architectures — driving up costs and complexity in the name of vendor roadmaps — QServe is a reminder that smart software still comes out on top.

As Dave accurately puts it:
”We’ve accepted that hardware has limitations, but there’s no reason to just lie down. You can be clever and work around it with software. Ingenuity always wins in the end.”

Why This Matters for Stelia

This is just one example of what happens when ingenuity reclaims center stage, refusing to be boxed in by hardware limitations. That principle resonates strongly with Stelia’s own purpose. At Stelia, we are committed to advancing artificial intelligence that is not just more powerful, but more purposeful — adaptive, transparent, and human-centered. While we continue to explore advanced quantisation, co-designed inference stacks, and systems-level optimisations like those demonstrated by QServe, our broader goal is to ensure these technologies serve humanity with responsibility, fairness and long-term impact. It is systems-led ideas like QServe that illustrate a path toward this goal — by proving that thoughtful engineering can break barriers and expand what AI is capable of.

References

[1] QServe Project — W4A8KV4 Quantization and System Co-design for Efficient LLM Serving: QServe: W4A8KV4 Quantization and System Co-design for Efficient LLM Serving
GitHub: GitHub — mit-han-lab/omniserve: [MLSys’25] QServe: W4A8KV4 Quantization and System Co-design for Efficient LLM Serving; [MLSys’25] LServe: Efficient Long-sequence LLM Serving with Unified Sparse Attention
Benchmarks Dataset: mit-han-lab/QServe-benchmarks · Datasets at Hugging Face
Lead Researcher: Song Han

Lab: MIT HAN Lab

Tobias Hooton, CEO and Founder, Stelia

A founder’s view on the next era of AI

Most industry observers still believe we’re in the early days of AI. They’re wrong.

The first era of artificial intelligence was a triumph of engineering. It was defined by breakthroughs in scale: GPUs got faster, models got bigger, projects like Stargate were launched. We built the infrastructure. We watched transformers reshape natural language, diffusion models redraw images and a totally new breed of AI-native products enter the mainstream. The energy of that period was electric. But it was also experimental, erratic and incomplete.

Today, the artificial tide is changing direction. We are entering the second era of AI and its shape is beginning to crystallise: not in academic labs or GPU clusters, but in enterprise conference rooms, legal firms, global supply chains, customer support teams, and the daily workflows of billions of people.

The first era of AI built the engine. The second era is about using it.

This moment marks a fundamental transition. And like any transition it comes with confusion and inertia. Many of the players who won in the first phase are now trying to force their way into continued relevance. But the centre of gravity has moved. The epicentre has shifted to application.

From Compute Wars to Value Wars

In the first era, value accrued to those who could build compute. Access to high-end GPUs became a competitive advantage. Entire business models formed around reselling compute and leasing hardware. (sounds similar to the early 2000s of bare metal compute…)

But here’s the uncomfortable truth: compute is not the moat. At best, it’s table stakes.

The AI industry is now saturated with infrastructure providers desperate to monetise their place in the stack. They speak the language of tokens and throughput, but these are engineering metrics, not user value. And while they focus on monetising bits and bandwidth, the real opportunity has evolved.

We are now witnessing a decoupling: those who control compute will not necessarily control the future. Value has shifted away from raw hardware and toward ease of use, access, and deployment at huge scale. AI is about making intelligence accessible: reliably, securely and in the right context.

The Second Phase of AI: Distribution at Scale

We are now in the “adoption era.” In this phase, the building is effectively done. The foundational breakthroughs required to build mass-market AI products have now occurred (research continues, but the core infrastructure exists). We’re no longer waiting on hardware; we’re waiting on distribution.

In this new landscape, the questions have changed: • The critical question: how well does it integrate into a 100,000-person enterprise? • Can it reduce legal review time by 40% and meet compliance requirements? • Can you scale it across 70 languages with consistent tone, brand, and results?

This is where the next generation of AI companies will win: in the trench warfare of deployment, serviceability and global distribution of intelligence.

And just as importantly, in usability.

Ease of use is not a feature; it’s the differentiator. Organisations want systems that augment their own, with minimal friction and maximum return. We’re seeing the rise of truly horizontal AI: core operational enhancements that drive measurable efficiency.

The era of experimentation is giving way to the era of execution.

Why the Old Guard Will Struggle

Who would be surprised by this prediction? The giants of Phase One who assume they’ll simply evolve into Phase Two. But history rarely works that way.

The incentives are misaligned. The largest compute providers (Nvidia partners, hyperscalers, GPU lessors) have built their business models around consumption primarily for a deeply technical niche audience, not value creation. They profit when you use more resources, not when you deliver better results. This creates a fundamental conflict: their growth is not your growth. Their scale is not your efficiency.

As enterprises begin to demand value instead of volume, these providers are being caught off-guard. They are structured to win hardware cycles, not workload loyalty.

We are already seeing cracks in their armour. Once-untouchable players are spinning up foundation model labs and launching APIs on top, hoping to capture enterprise spend. But APIs are not ecosystems, and tokens are not outcomes. Without deep understanding of user needs, without vertical integration, and without true control of the full stack, they risk becoming irrelevant middlemen in a fast-consolidating value chain.

While GPU providers remain a critical part of the AI ecosystem, powering the infrastructure that enables training and inference at scale, they have shifted from being the epicentre of innovation to a foundational layer in a much larger value stack. As the focus moves from raw compute to real-world utility, the differentiators are in how seamlessly intelligence can be deployed, adopted, and trusted. GPUs are essential but the future is decided elsewhere.

The Rise of Full-Stack Intelligence

In the second AI phase, the winners will be those who own the entire value chain: from silicon to sentiment. That means building systems, not just models.

This requires radical vertical integration and deep domain expertise. Control of the hardware. Custom foundation models. Scalable orchestration layers. And intuitive interfaces. All bound by a singular goal: make intelligence useable at scale.

This is where most AI startups fall short. They pick a piece of the puzzle (an LLM wrapper, a UX shell, an API filter) and try to build a business. Without full-stack control, they are at the mercy of upstream providers and downstream constraints. The result is fragmentation, not acceleration.

To thrive in this new world, you need to abstract the complexity. Hide the infrastructure. Deliver intelligence as a service: in the human sense, not the cloud computing sense.

And it must work across domains. Across cultures. Across languages. The next phase of AI is multicontextual. It’s about understanding what a user says, why they say it, when they say it, and how best to respond.

This is a systems challenge. And only a few players are positioned to solve it.

Why Stelia Already Has the Answer

At Stelia, we’ve been building for this moment for over 5 years, through our own deep research labs across multiple disciplines, we’ve centred ourselves around creating the technology stack truly distributed AI requires. A full-stack AI company: from GPU-level compute infrastructure, to in-house model training, to orchestration systems that serve hundreds of millions of users across sectors and geographies.

We didn’t chase hype cycles. We architected for scale. Our platforms were trained for the complexities of real-world deployment, not just benchmarks.

Where others scrambled to bolt on value, we built it into the foundation.

Stelia’s stack is an intelligence platform designed for mass adoption. From the kernel to the agentic interface, every component is optimised for enterprise use, regulatory clarity, and world-class experience.

This is Phase Two, purpose-built.

The artificial tides are changing and with them, the old paradigms are eroding. The future will be owned by those who can deliver the deepest impact.

At Stelia, we are the shift. And we’re just getting started.

In February 2025, European leaders announced InvestAI, a €200 billion initiative to build AI “gigafactories” across the continent. Commission President Ursula von der Leyen unveiled the programme on February 11th, describing it as creating “unprecedented capital through InvestAI for European AI gigafactories” and “a unique public-private partnership, akin to a CERN for AI.” France pledged €109 billion in private investment, whilst €20 billion was allocated specifically for four gigafactories housing approximately 100,000 last-generation AI chips. The rhetoric was stirring: technological sovereignty through infrastructure ownership, democratic oversight through geographical control, European values embedded in European silicon.

This represents one of the most expensive category errors in the history of technology policy.

The fundamental premise underlying Europe’s gigafactory strategy (that controlling physical infrastructure translates to meaningful governance of AI systems) reflects a profound misunderstanding of where power actually resides in algorithmic societies. This creates what we term the “sovereignty paradox”: the more sophisticated AI systems become, the less meaningful traditional territorial control mechanisms prove to be. Distributed AI systems resist the centralised monitoring and control that territorial governance assumes, whilst AI models exist as mathematical objects that can be instantly copied and deployed anywhere without degradation.

Recent events underscore this disconnect perfectly. When German authorities moved to block the DeepSeek AI app for unlawful data transfers to China, they targeted app store distribution rather than the underlying AI model itself. The mathematical weights powering DeepSeek’s capabilities remain unchanged and deployable through countless other channels, whilst enforcement focuses on controlling access points rather than algorithmic behaviour. This illustrates precisely how territorial governance mechanisms prove inadequate for governing distributed AI systems.

The Mathematical Absurdity of Territorial Control

The gigafactory approach rests on a fundamental category error about the nature of AI models themselves. Once training completes, an AI system exists as nothing more than a specific configuration of numerical parameters: billions or trillions of floating-point numbers arranged in precise mathematical relationships. These weights carry no inherent geographical identity, jurisdictional markers, or territorial constraints.

Consider the mathematical reality: a neural network trained on European data using European computational resources becomes functionally indistinguishable from an identical configuration of weights derived through any other process. The model’s “European-ness” exists only in its provenance metadata, not in its mathematical structure or functional characteristics. Two models with identical weights will produce identical outputs regardless of where they were trained, what data was used, or which regulations governed their creation.

This mathematical universality renders territorial control mechanisms absurd. A healthcare AI model trained in a European gigafactory using data from EU citizens under EU regulations can be perfectly replicated by simply copying its numerical weights. Once copied, those weights can be embedded in medical devices sold globally, integrated into healthcare systems operating under entirely different regulatory frameworks, or deployed on consumer smartphones without any technical mechanism for maintaining European governance.

The copying process itself reveals the deeper absurdity. Unlike physical goods that degrade through reproduction, mathematical objects can be perfectly duplicated infinitely without loss of fidelity. A single European-trained model can spawn millions of identical copies operating simultaneously across every jurisdiction on Earth, each mathematically indistinguishable from the “sovereign” original yet entirely beyond European control.

Distribution Trends and Governance Fragmentation

Whilst European policymakers focus on centralised training infrastructure, compelling technical and economic forces are driving AI inference towards distributed edge deployment. Real-time applications cannot tolerate the latency inherent in cloud-based inference: autonomous vehicles cannot wait 200 milliseconds for cloud processing during emergency braking; medical devices monitoring cardiac rhythm cannot pause for network connectivity during critical moments; industrial control systems require local processing capability that operates independently of network infrastructure and regulatory oversight.

Privacy requirements accelerate this distribution trend. Healthcare AI processing sensitive patient data locally avoids both regulatory complexity and data transmission risks. Financial algorithms running on consumer devices eliminate the legal and technical vulnerabilities associated with transmitting sensitive information to centralised processing facilities. Industrial systems processing proprietary operational data avoid competitive intelligence risks by keeping computation local.

The economics further reinforce distribution. Edge inference eliminates ongoing cloud computing costs, reduces bandwidth requirements, and enables offline operation. The global edge AI market, valued at $8.2 billion in 2024 and projected to reach $55.6 billion by 2030, reflects genuine economic value creation rather than regulatory compliance theatre.

These technical and economic forces create governance fragmentation that strikes at the heart of the gigafactory model. A single AI model trained in a European facility might simultaneously operate on German automobiles, French medical devices, Italian manufacturing equipment, and American smartphones, each subject to different regulatory regimes with no technical mechanism for unified governance. The temporal dimension intensifies this challenge: when AI models make split-second decisions on mobile devices, traditional accountability mechanisms prove inadequate for computational events occurring too quickly and too locally for meaningful oversight using territorial governance frameworks.

Early indications of ephemeral agent networks and autonomous agent-to-agent frameworks suggest even more challenging governance scenarios ahead. As AI systems evolve towards autonomous coordination patterns, these temporal and jurisdictional challenges will intensify exponentially.

The Sovereignty Paradox in Practice

At the architectural level, distributed systems resist the centralised monitoring and control mechanisms that territorial governance assumes. Edge AI operates through peer-to-peer networks, mesh topologies, and intermittently connected devices that collectively create computational capabilities without central coordination. Emerging agent-to-agent frameworks suggest even more autonomous coordination patterns ahead. These emergent system properties cannot be governed through control over any specific infrastructure component, including the facilities where models were originally trained.

The sovereignty paradox thus reveals a category error at the heart of European AI policy: attempting to govern inherently global, distributed, and stateless systems through territorial control mechanisms designed for physical, localised, and persistent objects. The €200 billion gigafactory investment doubles down on this error by strengthening the least relevant aspect of AI governance (control over training location) whilst ignoring the most consequential challenge (governance of distributed deployment).

Opportunity Cost and Alternative Approaches

Perhaps the most damaging aspect of the gigafactory approach lies not in what it attempts to achieve, but in what it prevents. The €200 billion investment represents resources that could fund governance innovations addressing the genuine challenges of democratic accountability in algorithmic societies.

Consider what €200 billion could accomplish if directed towards governance capability rather than training infrastructure: comprehensive legal frameworks for algorithmic transparency; technical infrastructure for citizen participation in AI development; international institutions for democratic AI governance; research and development for privacy-preserving collaborative AI; community capacity building for algorithmic accountability.

These investments would address the actual governance challenges posed by AI systems: ensuring algorithmic decisions can be explained and contested; enabling meaningful citizen participation in the development of systems that shape their lives; creating accountability mechanisms that work across jurisdictional boundaries; building technical infrastructure that embeds democratic values into system architecture.

Technical transparency mechanisms can embed governance properties directly into AI models through cryptographic attestation, making compliance verifiable regardless of deployment context. Rather than relying on territorial control over training facilities, these approaches create technical constraints that ensure algorithmic behaviour aligns with democratic values wherever deployment occurs. Participatory governance mechanisms can enable meaningful citizen input into algorithmic system design without requiring territorial control over computational infrastructure.

This misallocation occurs precisely when the window for governance innovation remains open. AI systems are still developing; technical standards remain fluid; international norms are still forming. European leadership in governance innovation could establish frameworks that other democracies adopt, creating global infrastructure for algorithmic accountability.

Beyond the Gigafactory Delusion

The gigafactory delusion represents more than misguided technology policy; it embodies a fundamental misunderstanding of power in algorithmic societies. By investing €200 billion in training infrastructure that provides no meaningful governance capability over distributed deployment, European leaders reveal dangerous confusion about where control actually resides in AI systems.

The technical realities of distributed AI systems, and the emerging prospect of ephemeral agent networks operating beyond traditional oversight mechanisms, demand nothing less than a fundamental reimagining of how democratic societies govern algorithmic power. The gigafactory approach represents exactly the wrong answer to exactly the right question.

Democratic societies deserve better: governance frameworks that enhance rather than undermine democratic accountability, investment priorities that strengthen rather than weaken citizen agency, and technology policies that serve democratic values rather than political theatre. Recognition of this error creates opportunity for approaches that could actually work if democratic societies have the intellectual courage to pursue them.

About Stelia

Stelia develops AI platforms that embed democratic accountability into technical architecture, proving that transparency and cooperation deliver superior outcomes to territorial control. Our work demonstrates that governance wisdom combined with technical authority can create AI systems that serve democratic values through design rather than oversight: exactly what distributed algorithmic societies require.

2025 isn’t even halfway through yet and it’s clear the shift from centralised to distributed inference is well underway. This represents perhaps the most significant AI execution trend of the year, and it’s at the heart of what Stelia will showcase as a Gold Sponsor at NVIDIA GTC Paris 2025 this June.

According to recent industry data, the global AI inference market is projected to reach $1.3 trillion by 2032, with distributed architectures driving remarkable cost reductions. Beyond technical tweaks this is new approach to how AI delivers business value.

Pilotitis

If you’ve been following the enterprise AI journey, you’re likely familiar with what industry analysts call pilotitis or the “prototype problem” — AI applications that work brilliantly in controlled environments but struggle when exposed to real-world conditions and scaling requirements.

“European enterprises have proven AI concepts but now face the challenge of reaching millions of users while maintaining performance,” explains Dan Scarbrough, Chief of Staff at Stelia. “The gap between today’s approaches and tomorrow’s AI-powered world creates bottlenecks that traditional methods cannot address.”

These represent fundamental limitations in latency, cost, and deployment when using centralised inference models:

• Latency issues when transferring data across distributed environments

• High operational costs from cross-region traffic and compute resources

• Data privacy and regulatory compliance challenges

• Inability to deploy effectively for real-time applications requiring millisecond responses

Lyra: The Distributed Intelligence Runtime

At NVIDIA GTC Paris 2025, Stelia will showcase how their Lyra distributed intelligence runtime transforms AI concepts into adaptive applications capable of reaching millions of users without boundaries.

Lyra addresses the fundamental challenges of centralised models through its agentic-led approach. The results are compelling:

• Latency reduction of up to 90%, from 150ms to as low as 10–15ms in latency-sensitive applications

• Bandwidth optimisation of 30–60% in content-heavy applications

• Agent orchestration with existing AI workflows and models

• Future-ready architecture through embedded, live-learning systems

What makes Lyra particularly powerful is its intelligence engine, which orchestrates autonomous agents and multi-agent systems-providing real-time intelligence as a core component of distributed AI applications.

The European AI Context

GTC Paris 2025 represents a unique convergence of AI development in the European context. Co-located with VivaTech (Europe’s largest startup and technology event), the conference will bring together developers, researchers, and business leaders from across the continent.

The timing couldn’t be more significant. According to recent data, 78% of organisations reported using AI in 2024, up from 55% in 2023. In Europe specifically, over 45% of EU enterprises intend to adopt edge-based AI by 2026.

NVIDIA CEO Jensen Huang’s live keynote on June 11 at the Dôme de Paris is expected to unveil exclusive European AI advancements, setting the stage for three days of intensive knowledge exchange and partnership development.

Collaborate to Win

Perhaps most compelling about Stelia’s presence at GTC Paris is their emphasis on design partnerships over pure technology provision. Their collaborative approach acknowledges that building transformative AI applications requires more than just powerful technology, demanding a shared journey from concept to global scale.

This design partnership model brings organisations from concept to global reach through a structured approach:

1. Understanding unique challenges and AI vision

2. Architecting applications on the Lyra intelligence engine

3. Deploying autonomous agents to enterprise customers or consumer audiences

4. Achieving global scale on a platform designed for distributed intelligence

For enterprises facing the reality that legacy approaches aren’t built for dynamic AI applications and that development velocity slows when trying to scale beyond initial success. This represents a compelling alternative to fragmented approaches.

Real-World Impact Across Sectors

The distributed intelligence revolution that Lyra enables is already transforming key sectors:

Media & Entertainment: Content delivery platforms require end-to-end latency below 50 milliseconds to maintain unified experiences. Distributed intelligence enables processing that reduces server load and bandwidth consumption by up to 40%.

Healthcare & Telemedicine: Real-time diagnostics, predictive alerts, and personalised treatments rely on fast, private AI processing. The market for AI-embedded healthcare wearables is projected to reach 350 million units by 2026.

Retail & E-commerce: AI-driven personalisation is expected to contribute $800 billion in new global retail revenue by 2025, with e-commerce retailers generating 10–30% of revenue from AI-driven suggestive selling.

Connect with Stelia at GTC Paris 2025

For organisations looking to scale their AI initiatives, Stelia’s team will be available throughout GTC Paris 2025, June 10–12 at the Paris Expo Porte de Versailles. To learn more about Stelia or to book a meeting during the event: visit Stelia, book a meeting or contact lyra@stelia.ai

Originally published at https://newsroom.stelia.ai on May 29, 2025.

Stelia, the distributed runtime for AI, is proud to announce their participation as a Gold Sponsor of NVIDIA GTC Paris 2025, colocated with VivaTech from June 10–12 at the Paris Expo Porte de Versailles.

As organisations move beyond AI experimentation to production deployment, they face fundamental challenges that traditional approaches cannot solve. At NVIDIA GTC Paris 2025, Stelia will showcase Lyra, the distributed AI runtime that transforms concepts into adaptive applications capable of reaching millions of users without technical barriers or complexity.

Lyra’s intelligence engine orchestrates all components of AI, providing borderless deployment and simple integration as a core component of scale-out AI applications. This architecture enables organisations to deliver consistent experiences regardless of location or device, eliminating the cost unpredictability and fragmentation that typically occurs when scaling beyond initial success.

Stelia experts will be available throughout the three-day event to demonstrate how Lyra’s distributed intelligence approach enables European enterprises and developers to build sophisticated AI applications that adapt, evolve, and create unprecedented value without rebuilding their underlying infrastructure.

Key Capabilities of Lyra

Lyra is purpose-built for intelligence at global scale, delivering:

• Borderless AI deployment — Simply distribute AI capabilities across regions and environments

• Adaptive runtime orchestration — Real-time personalisation and intelligence across global deployments

• Simplified integration — Connect with existing AI workflows and models without technical barriers

• Future-proof architecture — As AI evolves, your applications evolve with it

“European enterprises have proven the AI concept but now face the challenge of scaling to millions of users while maintaining performance,” said Dan Scarbrough, Chief of Staff at Stelia. “By partnering with forward-thinking organisations at GTC Paris, we’re demonstrating how our collaborative approach delivers tomorrow’s AI-powered experiences and measurable business outcomes, without boundaries.”

Meet the Stelia executive team, including Tobias Hooton CEO and Dan Scarbrough Chief of Staff, at NVIDIA GTC Paris 2025, June 10–12 at the Paris Expo Porte de Versailles.

For more information about Stelia and how we can benefit your organisation, please visit Stelia or contact lyra@stelia.ai

About Stelia:

Stelia delivers intelligence at global scale through Lyra, the distributed AI runtime. Lyra transforms AI concepts into adaptive applications, reaching millions of users without barriers or complexity. Lyra orchestrates AI components for borderless deployment and simple integration, driving experiences that adapt, evolve, and unlock unprecedented value for innovators ready to harness AI’s full potential. Stelia delivers measurable business outcomes, without boundaries.

About NVIDIA GTC Paris 2025

NVIDIA GTC Paris 2025 brings together Europe’s most innovative AI developers, engineers, and researchers to explore advancements in AI, accelerated computing, and related fields. Held from June 10–12 at the Paris Expo Porte de Versailles, the conference features a keynote by NVIDIA CEO Jensen Huang on June 11 at the Dôme de Paris. The event emphasises real-world AI applications, agentic AI, generative AI, and cloud infrastructure, fostering collaboration between NVIDIA experts and the European tech community.

About VivaTech 2025

VivaTech is Europe’s largest startup and technology event, held annually in Paris, France. The 2025 edition, running from June 11–14, will welcome over 165,000 attendees from more than 160 countries, featuring 2,800+ exhibitors and 13,500+ startups. The event focuses on technological innovation, with special emphasis on AI advancement, sustainability, and cross-border collaboration.

Originally published at https://newsroom.stelia.ai on May 27, 2025.

Middle East Emerges as Key Hub for Distributed Intelligence

The United States has signed landmark agreements with the United Arab Emirates and Saudi Arabia that will fundamentally reshape the global AI landscape. These deals unlock a trillion-dollar capital pool while positioning the Middle East as a crucial node in the emerging distributed intelligence ecosystem.

Reshaping the Global AI Landscape

According to SemiAnalysis, “The new U.S. pacts with the UAE and Saudi Arabia reshape the AI landscape on three fronts: Macro, Geopolitical, and Infrastructure.” By sidestepping previous export-control frameworks, Washington has opened significant capital flows for AI infrastructure development.

The UAE agreement centers around Abu Dhabi’s G42, securing an import quota of 500,000 Nvidia top-tier chips annually. The deal includes a planned 5GW AI datacenter campus, with the first 1GW phase already under construction, utilizing solar, gas, and nuclear power sources.

Trillion-Dollar Capital Mobilization

Saudi Arabia’s parallel $600 billion economic package includes DataVolt’s $20 billion investment in U.S. datacenters, alongside major commitments from Oracle, Google, and other tech giants. The newly formed Saudi AI firm HUMAIN will deploy approximately $10 billion in AMD systems and an equal investment in Nvidia hardware.

These agreements address critical infrastructure challenges. As SemiAnalysis notes, “Energy-rich Gulf nations join the roster of trusted partners just as U.S. data-center grids hit their physical limits. Europe could have relieved the bottleneck but stumbled on power shortages and slow permitting.”

From Centralized to Distributed Intelligence

“We’re moving from centralized AI models to distributed intelligence networks that can scale globally,” observes Tobias Hooton, CEO of Stelia. “These agreements go way beyond building more datacenters towards a new understanding of how intelligence is distributed and activated across geography.”

The deals also reshape the competitive landscape with China. As the Department of Commerce issued guidance declaring the use of Huawei Ascend chips an export control violation, these partnerships effectively extend U.S. technological influence in the region.

Security and Governance Frameworks

Security considerations remain important. The White House AI Czar David Sacks has expressed confidence in management strategies, noting that physical verification is straightforward: “all one would have to do is send someone to a datacenter and count the server racks to make sure the chips are still there.”

Industry Leaders Convene in Dubai

As the industry processes these developments, key stakeholders will gather at the Digital Infrastructure Investment Forum in Dubai. Stelia CEO Tobias Hooton will participate in a panel discussion on Monday, May 19th at 2:15pm at Madinat Jumeirah, alongside representatives from TAWAL, Maerifa Solutions, and Khazna Data Centers. The panel will explore capital flows, connectivity, scaling demands, and investment models.

Activating the Platform Intelligence Layer

“While much focus has been on building physical infrastructure, the real transformation will come from how we activate these resources,” says Hooton. “The most innovative companies are focusing on delivering business outcomes from platform intelligence that can activate, distribute, and scale in-market.”

These historic agreements create powerful new vectors for innovation while reshaping geopolitical relationships. The trillion-dollar question becomes how effectively this distributed intelligence architecture can address our most pressing challenges at global scale.

Originally published at https://newsroom.stelia.ai on May 19, 2025.

Distributed intelligence provides the foundational architecture for scaling enterprise AI, delivering the latency, cost, and performance advantages discussed in our previous analysis. However, successful implementation requires navigating significant organisational and technical challenges.

Our exploration of enterprise AI scaling has revealed both the implementation challenges facing organizations and how distributed intelligence provides the architectural foundation for success. In this final installment, we examine how Stelia’s agent-led platform converts these concepts into operational reality.

The Build vs. Buy Decision

The demands for scaled AI are substantial, with system complexity challenges cited by 66% of executives in the KPMG survey.

Organisations face two pathways:

1. Internal development — Requiring substantial capital investment and specialised expertise

2. Platform partnership — Leveraging purpose-built solutions from specialised providers

While both approaches can work, the second option significantly accelerates time-to-value. Stelia’s distributed intelligence platform removes the need for complex management, enabling enterprises to scale AI capabilities efficiently without extensive capital outlays or specialised technical teams.

Orchestration Through Abstraction

The key innovation driving Stelia’s approach is resource abstraction — a software layer that shields business users from underlying complexity while dynamically optimising system performance. This orchestration layer, powered by Stelia, delivers four critical advantages:

1. Simplified Management: Abstracts complex configurations, reducing the expertise needed for system orchestration and addressing the skills gaps (51%) identified in the KPMG survey.

2. Dynamic Scalability: Automatically allocates resources to meet fluctuating AI workload demands.

3. Cost Optimisation: Continuously evaluates and optimises resource usage for maximum efficiency.

4. Deployment Flexibility: Supports hybrid deployment models that address data sovereignty requirements (a priority for 54% of organisations) while minimising latency (56% prioritise edge computing).

From Technical Complexity to Business Value

Managing distributed AI systems traditionally requires specialised expertise in orchestration and scalability — a significant barrier given the technical skills gaps (51%) identified in the KPMG survey.

The Stelia advantage lies in transforming this complexity into intuitive workflows. Rather than requiring teams to manage intricate system configurations, Stelia’s agent-led approach provides natural language interfaces that enable non-specialists to deploy and manage sophisticated AI capabilities.

This approach extends beyond system management to several optimisation techniques that industry experts identify as critical for AI efficiency:

1. Model Optimisation: Stelia’s intelligent orchestration applies approaches that can reduce model complexity while preserving accuracy.

2. Continuous Monitoring: The platform tracks utilisation in real-time, identifying bottlenecks before they impact performance.

3. Automated Resource Management: Dynamically adjusts resources based on demand patterns, ensuring optimal performance and cost efficiency.

These capabilities directly address the data consistency and quality concerns (64%) identified as major barriers to AI adoption. By automating synchronisation and validation across distributed nodes, Stelia ensures reliable model performance even in real-time applications.

From Possibility to Production

Enterprise AI stands at an inflection point. Organisations have validated AI’s business value through pilots but struggle to scale these initiatives securely and cost-effectively across operations.

Distributed intelligence, powered by Stelia’s adaptive orchestration layer, provides the foundation for this transition — transforming AI from experimental projects to operational capabilities. By simplifying management, enhancing scalability, and optimising costs, Stelia enables organisations to overcome the financial, technical, and regulatory barriers to enterprise-wide AI deployment.

The KPMG data makes clear that competitive advantage in AI will not come from incremental experimentation but from operational excellence at scale. Organisations that can bridge the gap between pilot success and enterprise deployment will capture disproportionate value in the AI economy.

The future of AI-driven enterprises will be defined not by those who experiment most broadly, but by those who execute most precisely.

Originally published at https://newsroom.stelia.ai on May 9, 2025.

To overcome the scaling challenges outlined in Part 1, leading enterprises are embracing distributed intelligence — a fundamentally different approach to AI deployment that transcends traditional limitations.

In our previous analysis, we examined the KPMG data revealing a critical execution gap in enterprise AI, with 65% of organizations stuck in pilot phases. Now we explore how Stelia’s distributed intelligence architecture fundamentally transforms this AI scaling equation.

The Hidden Cost of AI Inference

While much attention focuses on model training, recent research reveals that over 80% of computational demand in AI actually comes from inference tasks — the everyday predictions and decisions that deliver business value. Inefficient inference operations lead to wasted resources that diminish ROI and can ultimately discourage broader AI adoption.

Organisations must recognise that scaling AI isn’t merely about performance but about sustainable economics. Stelia’s distributed intelligence platform addresses this precise challenge.

Beyond Single-Node Computing

Conventional AI deployments concentrate computing resources in centralised locations — creating bottlenecks, latency issues, and prohibitive scaling costs. Distributed intelligence takes a different approach by orchestrating AI workloads across optimal computational resources based on performance, cost, and compliance requirements.

This orchestration layer enables continuous, high-throughput processing and real-time decision-making by intelligently directing AI models across multiple nodes. The approach addresses fundamental deployment barriers by:

Ensuring data sovereignty compliance — Processing data where regulations require while maintaining unified workflows

Reducing latency through proximity — Critical for real-time applications in customer experience and operational analytics

Optimising resource utilisation — Directing workloads to the most efficient processing environment for each specific task

The Adaptive Advantage

Stelia’s distributed intelligence approach provides significant performance advantages across critical metrics. Independent benchmarks have demonstrated that similar distributed approaches can improve:

• Latency — Delivering superior end-user experiences by minimising the time from prompt submission to result delivery

• Throughput — Processing substantially more AI inferences within the same timeframe

• Cost efficiency — Translating directly to improved ROI and freed resources for innovation

These aren’t theoretical projections but reflect the real-world advantages of intelligent workload orchestration. The approach directly addresses the execution challenges revealed in the KPMG survey, enabling enterprises to accelerate AI deployment beyond pilot phases.

Direct Business Impact

The business case for distributed intelligence becomes increasingly compelling as enterprises seek to scale their AI initiatives:

1. Cost Efficiency: Stelia’s distributed orchestration delivers up to 86.4% savings compared to traditional cloud approaches, addressing the financial barriers to scaled AI deployment identified in the KPMG survey.

2. Data Security & Compliance: By processing sensitive information locally while maintaining unified workflows, Stelia mitigates the risk management concerns cited by 82% of executives.

3. Operational Resilience: Stelia’s intelligent orchestration layer ensures AI systems can operate efficiently across diverse workflows, critical for supporting the expanding use cases in recruitment (26%), call centers (61%), and data analysis (78%).

4. Simplified Deployment: Stelia’s adaptive intelligence layer abstracts away infrastructure complexity, addressing the technical skills gaps (51%) that hinder adoption.

The market validates this approach. The AI inference segment is projected to grow from $106.15 billion in 2025 to $254.98 billion by 2030 (19.2% CAGR), driven by enterprises seeking real-time insights to maintain competitive advantage in data-intensive sectors.

Having examined the technical foundation of distributed intelligence, in our final installment we address the practical implementation challenges organizations face and how Stelia’s agentic approach transforms complex orchestration into intuitive business workflows.

Originally published at https://newsroom.stelia.ai on May 6, 2025.

Stelia was on the ground at POSSIBLE 2025, meeting face-to-face with brand and media leaders grappling with how to make AI work at scale. Energy was high — but urgency was higher. With 33% of attendees representing brand leadership (a sharp rise from 21% in 2023) and over 2,700 meetings booked through the app, this wasn’t a trade show. It was a high-stakes summit on the future of media, content, and technology. And AI was at the center of nearly every conversation.

But despite all the excitement around agent-led orchestration and next-gen creative ops, one message cut through: Media & Entertainment isn’t struggling with ambition — it’s stuck in execution.

We heard it repeatedly in meetings and on the floor:

“We’re stuck in pilots.”

This frustration wasn’t theoretical. Companies have AI initiatives in motion — but they’re moving slowly. Many are siloed, reliant on internal engineering, and bound by the complexity of cloud costs and rollout delays. One executive summed it up plainly:

“We need to personalize globally without killing the team.”

That tension — between scale and sustainability — is now defining the industry’s relationship with AI. Creative leads want autonomy. Ops teams want velocity. CMOs want content out the door faster than ever. But the stack beneath them -piecemeal tooling, experimentation-heavy workflows, and operational overhead -isn’t keeping up.

And it’s leading to hesitation:

“Cloud costs and rollout delays are making us hesitate.”

These weren’t isolated complaints. They surfaced across panels like “The Proximity Shift: Agentic AI and the Future of All Brands” and “The Agentic Revolution: Reimagining Visual Content Marketing with AI.”

The window for “trying AI” has closed. If you’re not scaling, you’re stalling.

That’s where our presence at POSSIBLE was most felt — not in promoting a product, but in sharing a model that directly addresses the scale gap. The goal is simple: get platforms out of pilot mode and into production, faster.

At Stelia, we’ve designed for exactly these M&E constraints:

• Flat-fee platform models that eliminate budget ambiguity and align with how content teams actually operate

• Fast-deployment architecture that avoids month-long engineering timelines and cloud integration purgatory.

• Agent-led orchestration that liberates creative teams from manual dependencies and accelerates global personalization without team burnout.

At POSSIBLE, what connected wasn’t concept — it was clarity. A model that maps directly to the pain points leaders are living. When we described a platform approach that replaces engineering-heavy workflows with intelligent orchestration — and does it at predictable, non-extractive pricing — the response was immediate: “This is what we’ve been waiting for.”

For Media & Entertainment teams pushing toward AI maturity, it’s time to move from ambition to orchestration. From pilots to platforms. From bottlenecks to flow.

Originally published at https://newsroom.stelia.ai on May 2, 2025.

Artificial intelligence has moved beyond hype to become a critical competitive differentiator. The latest KPMG AI Quarterly Pulse Survey (Q1 2025) confirms what forward-thinking organisations already know: 93% of enterprise leaders report that generative AI investments have strengthened their competitive position. Yet a striking paradox exists. Despite widespread experimentation and proven business value, only 11% of organisations have fully deployed AI agents, while 65% remain stuck in pilot phases. This execution gap represents the central challenge facing enterprise AI initiatives today

The Shift from Experiment to Execution

The data shows clear momentum in AI adoption across critical business functions:

• 78% are leveraging AI for data analysis (up from 70%)

• 66% for administrative duties (up from 27%)

• 61% for call centers (up from 16%)

• 26% for recruitment (up from 15%)

These impressive growth rates reflect a maturing understanding of AI’s potential. Organisations increasingly view AI as an augmentation tool rather than a replacement technology — 76% believe it will automate tasks without replacing roles, while 69% see it empowering high performers.

The Technical Inflection Point

As AI initiatives move from conceptual to operational, leadership dynamics are shifting dramatically. CIO involvement in AI strategies has surged from 31% to 86%, while CEO involvement has decreased from 34% to 8%. This technical handoff reflects the evolution from strategic vision to execution reality — precisely where many organisations encounter significant barriers

Investment vs. Implementation

Organisations are backing their AI ambitions with substantial capital, planning to invest $114 million in generative AI over the coming year (up from $89 million). With 97% reporting measurable profitability, the business case is clear. Yet executives consistently identify five critical barriers preventing scaled deployment:

1. Risk Management (82%): Cybersecurity vulnerabilities, compliance issues, and operational risks

2. Data Quality (64%): Inconsistent, siloed data undermining model accuracy

3. Trust (35%): Building workforce confidence amidst concerns about job displacement

4. Technical Complexity: System architecture challenges (66%), rapid technology evolution (56%), and skills gaps (51%)

5. Cost Efficiency: Substantial complexity and financial exposure in scaling AI operations

The paradox is clear: enterprises have validated AI’s business value through pilots but lack the orchestration layer needed to scale these initiatives securely, reliably, and cost-effectively across global operations.

With the enterprise AI scaling challenges clearly established, our next article explores how Stelia’s distributed intelligence approach creates the architectural foundation necessary to bridge the gap between AI experimentation and enterprise-wide deployment.

Originally published at https://newsroom.stelia.ai on April 30, 2025.