The State of AI 2026

How much did frontier models actually improve?

Frontier AI improved faster in 2025-2026 than the benchmarks tracking it could absorb. The Stanford HAI AI Index 2026 reports that models gained 30 percentage points in a single year on Humanity's Last Exam, an evaluation deliberately designed to be difficult for AI and favorable to human experts. Gains were largest on the tests that were well below human level only a few years ago: on GPQA Diamond, PhD-level "Google-proof" science questions, Epoch AI records Grok 4 at 87% (as of July 2025), roughly 17 points above the 69.7% scored by the PhD-level human experts OpenAI recruited to set the baseline.

The most important structural change is convergence at the top. As of March 2026, four to six labs cluster within 25 Elo points on the human-voted LMArena Leaderboard: Anthropic (1,503), xAI (1,495), Google (1,494), OpenAI (1,481), Alibaba (1,449), and DeepSeek (1,424) (Stanford HAI AI Index 2026). When capability stops being a clear differentiator, competition shifts to cost, reliability, latency, and domain fit, precisely the pattern the Index documents.

That convergence is now geographic as well as competitive. The US-China model gap has effectively closed: DeepSeek-R1 briefly matched the top US model in February 2025, and by March 2026 the leading US model led by just 2.7%, a gap that stayed in the single digits all year (Stanford HAI AI Index 2026).

Are AI benchmarks saturating?

Yes, and faster than their designers intended. On SWE-bench Verified, where models must resolve real GitHub issues, performance rose from roughly 60% in 2024 to close to 100% in 2025 (Stanford HAI AI Index 2026). GPQA Diamond, MMMU, and AIME have all been pushed to or past their human baselines. Evaluations built to stay relevant for years are now saturating in months, compressing the window in which any single benchmark tracks real progress.

Saturation has exposed a measurement crisis. The AI Index 2026 finds error rates of up to 42% on widely used evaluations, ranging from 2% flawed questions on MMLU Math to 42% on GSM8K, and separate research suggests Arena standing may partly reflect adaptation to the platform rather than general capability. In response, the Index rescales each benchmark so that 105% means a model performs 5% better than the human baseline, an admission that raw scores near 100% no longer carry clean signal.

Not everything has fallen. FrontierMath remains the frontier that resists saturation: at its late-2024 launch, every model tested solved under 2% of its research-grade problems (Epoch AI), and its hardest tier still separates the best reasoning systems from the rest. FrontierMath is the current example of a benchmark hard enough to keep measuring what models genuinely cannot yet do.

What does the 2026 model landscape look like — closed vs open weights?

Six families define the frontier: OpenAI's GPT-5 line, Anthropic's Claude, Google's Gemini, xAI's Grok, and the open-weight challengers led by DeepSeek and Alibaba's Qwen, with Moonshot's Kimi and Meta's Llama filling out the ecosystem. The AI Index 2026 timeline captures the pace: GPT-5.1 (Nov 2025) reached about 76.3% on SWE-bench Verified versus roughly 72.8% for GPT-5; Gemini 2.5 Pro (Mar 2025) shipped a 1M-token context and hit #1 on LMArena; and Claude Sonnet 4.5 (Sep 2025) posted 61.4% on OSWorld computer-use and 77.2%+ on SWE-bench Verified.

The open-versus-closed gap reopened after nearly closing. As of March 2026, the top closed model led the top open model by 3.3%, up from just 0.5% in August 2024, and six of the ten highest-ranked Arena models are now closed (Stanford HAI AI Index 2026). Open weights match proprietary models on a growing list of everyday tasks but trail on the broadest, hardest reasoning, keeping a thin but real closed-model premium.

The economic stakes of open weights became undeniable with DeepSeek-R1 (Jan 2025), whose reinforcement-learning training method (GRPO) delivered frontier-adjacent reasoning at far lower cost. Its release wiped over one trillion dollars off US technology stocks as investors reassessed whether cheaper training would erode incumbents' moats (Stanford HAI AI Index 2026).

Did reasoning models and agents deliver in 2026?

Reasoning models were the defining architecture of the period. Systems that spend inference-time compute "thinking" before answering drove the gains on math, science, and coding: Google's Gemini Deep Think scored 35 points, a gold medal, at the 2025 International Mathematical Olympiad, working end-to-end in natural language inside the 4.5-hour limit, up from the 28-point silver achieved in 2024 (Stanford HAI AI Index 2026).

Agents advanced from answering questions to completing multi-step tasks. On OSWorld, which tests agents on real computer tasks across operating systems, accuracy rose from roughly 12% to 66.3% in a year, within six points of human performance (Stanford HAI AI Index 2026). Tooling matured alongside capability: Claude Sonnet 4.5 shipped checkpoints, memory editing, and the Claude Agent SDK to let developers build long-running autonomous workflows.

The caveat is reliability. Agents still fail roughly one in three attempts on structured benchmarks, which is the difference between an impressive demo and a system you can trust unsupervised in production. The unlock in 2026 was not perfect agents but agents good enough to keep a human in the loop rather than doing the whole task by hand.

What can AI still not do?

Capability in 2026 is jagged, not uniform. Researchers use the term "jagged intelligence" to describe systems that win a math-olympiad gold medal yet fail at tasks a child handles: on ClockBench, the top model read analog clocks correctly just 50.6% of the time, against 90.1% for humans (Stanford HAI AI Index 2026). High scores on abstract reasoning do not guarantee competence on mundane, physical, or perceptual tasks.

The physical world remains the hardest frontier. Robots succeed on only about 12% of real household tasks even as they reach 89.4% on software-based RLBench simulations, and the gap between controlled labs and unpredictable homes stays wide (Stanford HAI AI Index 2026). Autonomous vehicles are the exception, with Waymo running roughly 450,000 weekly trips across five US cities, but only in favorable conditions with remote human fallback.

Professional reliability is the other open problem. In tax, mortgage processing, corporate finance, and legal reasoning, model performance ranges from 60% to 90%, and the top 15 models are separated by as little as 3 percentage points, so no system yet clears the accuracy bar these high-stakes domains demand (Stanford HAI AI Index 2026). For finance and law, 90% is not good enough, and that last stretch is where 2026's models still fall short.

How fast is AI compute actually growing?

Fast, and without a visible ceiling. Epoch AI finds that the training compute of frontier AI models has grown 4–5× per year since 2010, with frontier language models running closer to 5× per year since 2020 (Epoch AI, 2025). Compounded, that means the compute behind a leading model roughly doubles every six to ten months — an order of magnitude every two years.

This growth is not a single lever but three multiplying together: larger GPU clusters, longer training runs, and better hardware (Epoch AI, 2025). The shift toward reasoning models has added a fourth demand centre — inference-time compute — but has not slowed training scaling. Epoch expects the 4–5× annual pace to hold near-term, which is why every downstream number in this chapter, from silicon orders to megawatts, is moving on the same exponential.

What does a 2026 AI data centre look like?

It looks like a power plant with servers attached. The reference point is xAI's Colossus 2 in Memphis, Tennessee: Epoch AI estimates it holds about 1.11 million H100-equivalents of compute drawing 946 MW of IT power, built on Nvidia B200 and B300 chips at a capital cost near $36 billion (Epoch AI Data Centers, 2026). Construction began in March 2025; to energise the site quickly, xAI installed natural-gas turbines nearby rather than wait for grid interconnection.

Colossus 2 is the leading edge of a broad move into the gigawatt class. Individual training sites have crossed a threshold where their power draw rivals a mid-sized city, and where the scarce inputs are no longer chips alone but land, cooling, transformers, and grid connections. The build-out is also concentrated: five hyperscalers — Amazon, Google, Meta, Microsoft and Oracle — collectively controlled an estimated 71% of the world's cumulative AI compute by Q4 2025, up from 63% in early 2024 (Epoch AI, 2025).

Who controls the supply chain — and where are the chokepoints?

A handful of firms sit at every narrow point. Nvidia supplies an estimated 80–90% of AI accelerators by revenue, with industry estimates putting its 2025 share near 85–86%; its data-centre revenue reached $115.2 billion in fiscal 2025, more than double the prior year (Nvidia disclosure; Silicon Analysts, 2025). AMD is a distant second at roughly 6%, with hyperscaler custom silicon — Google TPUs, Amazon Trainium — the main structural challenger.

Below Nvidia, the chokepoints tighten. TSMC fabricates essentially all leading-edge AI accelerators, and its advanced-packaging (CoWoS) capacity is the true bottleneck on how many can ship. TSMC's most advanced nodes depend on extreme-ultraviolet lithography, for which ASML of the Netherlands is the sole supplier worldwide. And every top accelerator needs high-bandwidth memory (HBM), made by just three companies — SK Hynix, Samsung and Micron — with SK Hynix holding roughly half the market (TrendForce, 2025). Any one of these — CoWoS, EUV, HBM — can gate the entire industry.

Why is compute so concentrated in the United States?

Because capital, cloud demand and chip access all point the same way. The United States hosts about three-quarters of global GPU-cluster performance in Epoch AI's dataset as of May 2025, with China second at roughly 15% (Epoch AI, 2025). The gap is reinforced by US export controls that restrict China's access to Nvidia's most advanced chips and to EUV lithography.

This produces the US–China paradox of AI compute. China leads the world in electricity generation and can build data-centre shells and power faster than almost anyone — yet it trails badly in installed frontier compute because the scarce input is advanced accelerators, not concrete or megawatts. The US has the opposite problem: abundant chips but an increasingly strained grid. In short, America is compute-rich and power-constrained; China is power-rich and compute-constrained. Ownership has also shifted decisively to industry, which held about 80% of the compute in Epoch's dataset by 2025, up from roughly 40% in 2019 (Epoch AI, 2025).

Can the grid keep up with AI's energy demand?

This is now the binding constraint on AI, and the honest answer is: only with strain. The IEA estimates data centres consumed about 415 TWh in 2024 — roughly 1.5% of global electricity — and projects that doubling to about 945 TWh by 2030, just under 3% of world consumption, with AI-driven accelerated servers growing about 30% per year (IEA, Energy and AI, 2025). In the US alone, data-centre demand is projected to rise by up to 240 TWh, a 130% increase over 2024 levels.

The capital confirms the scale. The largest US cloud and AI providers — Microsoft, Alphabet, Amazon, Meta and Oracle — have signalled combined 2026 capital expenditure in the range of roughly $660–725 billion, most of it on data centres, chips and power infrastructure, nearly doubling 2025's already-record spend of over $378 billion (company disclosures; Statista; Futurum, 2026). The bottleneck is shifting visibly from silicon to substations: interconnection queues, transformer shortages and local grid capacity now decide where compute can physically be built, which is why operators like xAI are self-generating power on-site rather than waiting for the grid.

How big is the AI market — and what do the long-range forecasts say?

The global artificial-intelligence market was valued at roughly $390.9 billion in 2025 and is projected to reach about $3.5 trillion by 2033, a compound annual growth rate near 30.6% for 2026–2033, according to Grand View Research. Forecasts vary by scope and methodology — an earlier Grand View estimate put the market at $1,811.75 billion by 2030 at a 36.6% CAGR — so the precise number depends on which firm and which definition you cite. What is consistent across sources is the direction and the slope: double-digit annual growth off an already large base.

The macro-impact forecasts are larger still, because they measure AI's spillover into GDP rather than software revenue. PwC's "Sizing the Prize" analysis estimates AI could add up to $15.7 trillion to the global economy by 2030 — about a 14% uplift versus a no-AI baseline — split between productivity gains and increased consumer demand. McKinsey estimates that generative AI specifically could add the equivalent of $2.6 trillion to $4.4 trillion annually across the 63 use cases it analyzed, lifting the impact of all AI by 15–40%. These are modeled potential ceilings, not realized results, and should be read as scenario forecasts rather than booked value.

How much private capital went into AI in 2025?

Private AI investment set a record in 2025, and it is heavily American. US private AI investment reached $285.9 billion, versus $12.4 billion in China — a gap of about 23 times, per Stanford HAI's 2026 AI Index. Generative AI drove the surge: it grew more than 200% year over year and captured nearly half of all private AI funding, and in generative AI specifically US investment exceeded the combined total of China and Europe by a wide margin.

AI is also absorbing an outsized share of the broader venture market. Private investment into AI grew 127.5% and, on Stanford HAI's accounting, private capital now represents about 60% of total AI investment globally. In US venture terms, AI-related companies captured a record portion of all dollars deployed in 2025 (PitchBook), concentrating capital that in prior cycles would have spread across a wider set of sectors.

One caveat on the US-China gap: the $12.4 billion figure counts private investment only. China channels large sums through government guidance funds — an estimated $184 billion directed toward AI companies between 2000 and 2023 — so private-investment totals alone understate China's full commitment (Stanford HAI).

What are the leading AI labs actually worth?

Valuations for the frontier labs are private-market marks set in funding rounds and secondary sales, not public prices, so they move fast and carry wide reported ranges. OpenAI was valued at roughly $500 billion in an October 2025 employee secondary share sale, up from $300 billion earlier in 2025, with higher figures reported in subsequent 2026 financing discussions. Anthropic was valued at about $183 billion in its September 2025 round, with reports of talks in early 2026 pointing toward $350 billion or more. Elon Musk's xAI was valued at roughly $50 billion in a 2024 raise, with substantially higher figures reported in 2025–2026 financing and restructuring talks. All of these are dated snapshots; treat any single figure as a marker from a specific round, not a settled price.

The structural story matters more than any one mark: a very small number of US labs command the majority of frontier-model capital. That concentration is itself a finding — it shapes competition, compute access, and where returns (if they materialize) accrue.

How much are Big Tech and the hyperscalers spending on AI?

Beyond startup funding, the largest US technology companies are financing AI directly through capital expenditure on data centers, chips, and power. The four largest US hyperscalers — Microsoft, Alphabet, Amazon, and Meta — guided to well over $300 billion in combined capital expenditure for 2025, up sharply from roughly $230 billion in 2024, according to their own earnings disclosures. Much of this spend is directed at AI infrastructure: GPUs, custom accelerators, and the electricity and cooling to run them.

This corporate capex is a different risk category than venture funding. It is deployed by profitable public companies against their own balance sheets, which makes it more durable than startup rounds — but it also front-loads enormous fixed costs against AI revenue that is still ramping, which is precisely the tension underneath the bubble debate.

Is AI a bubble? What the valuation-versus-revenue data shows

The honest answer is that the data supports both readings, and reasonable analysts disagree. On the caution side: frontier labs trade at high multiples of revenue. OpenAI reportedly reached roughly $13 billion in annualized revenue by mid-2025 (up from about $3.7 billion in 2024), which against a ~$500 billion valuation implies a revenue multiple far above mature software norms; Anthropic's reported run-rate, in the several-billion-dollar range over the same period, sits against a similarly outsized mark. On the demand side, an MIT report ("State of AI in Business," August 2025) found that roughly 95% of enterprise generative-AI pilots had produced no measurable return, suggesting a gap between deployment and realized value.

On the other side: revenue is growing at rates rarely seen at this scale, adoption is broad and rising, and the largest buyers of AI infrastructure are cash-generative incumbents rather than speculative newcomers — a materially different profile from the 1999–2000 dot-com cohort. The concentration of capital in a few US labs also means that a repricing of those specific marks would not automatically imply a broad market collapse.

This is not investment advice. The purpose here is to state the measurable facts — record investment, high private-market valuations, high revenue multiples, rapid revenue growth, and uneven realized returns — and let readers weigh them. No named primary source in this chapter declares AI definitively a bubble or definitively not one.

How many people actually use AI in 2026?

Consumer adoption has moved from novelty to habit. OpenAI's Sam Altman confirmed ChatGPT passed 800 million weekly active users in September 2025 — close to 10% of the world's adults — and the product was reported at roughly 900 million weekly users by February 2026 (OpenAI, via TechCrunch, February 2026). That is the fastest ramp of any consumer software in history: ChatGPT went from 400 million weekly users in February 2025 to 700 million by July 2025.

The category is bigger than any single product. Stanford's AI Index 2026 estimates generative AI reached 53% adoption among the US population within three years of launch — faster than either the personal computer or the internet at the same stage. Stanford also puts the annual consumer value of generative-AI tools to US users at roughly $172 billion by early 2026, with the median value per user tripling between 2025 and 2026. AI has become an everyday utility, not an experiment.

How many companies have adopted AI?

Enterprise adoption is now near-universal on paper. McKinsey's State of AI 2025 survey found 88% of organizations use AI in at least one business function, up from 78% a year earlier, and 71% regularly use generative AI in at least one function. Adoption is broadest in marketing and sales, product and service development, and IT.

Official statistics that count firms rather than survey respondents land lower, because they include the whole economy — small firms included. In the EU, Eurostat reported that 20.0% of enterprises with 10 or more employees used AI in 2025, up sharply from 13.5% in 2024 (Eurostat, December 2025). US federal data tells a similar story of a wide long tail: the Federal Reserve's tracking of Census Bureau data shows AI use among all US firms still in the low double digits, even as the largest companies race ahead. The headline: big organizations have adopted AI; the median small business is only starting.

Which industries lead — and what do the headline numbers say?

Finance is a front-runner. In banking, an EY-Parthenon survey found 77% of banks had actively launched or soft-launched generative-AI applications, up from 61% in 2023, and a Temenos survey of 420 global financial-services firms found 75% exploring generative-AI deployment. Adoption is driven by fraud detection, document processing and customer service.

Healthcare adoption is visible in the regulatory record: the US FDA had authorized roughly 1,450 AI/ML-enabled medical devices cumulatively by the end of 2025 — with a record 295 cleared in 2025 alone — the large majority in radiology (FDA device list, 2025). Legal, retail, manufacturing and education are further behind but moving: legal work is being reshaped by document-review and drafting copilots; retail leans on demand forecasting and personalization; manufacturing on predictive maintenance and quality inspection; and in education, generative AI has become the default study aid for students, forcing a rapid rethink of assessment. The pattern across sectors is consistent — deep pilots in a few high-value tasks, shallow penetration everywhere else.

Is agentic AI actually deployed yet?

Agentic AI — systems that plan and execute multi-step tasks with limited human oversight — was the dominant enterprise narrative of 2025-2026, but real deployment lags the hype. Stanford's AI Index 2026 notes that agentic deployment remains in the single digits across most business functions, even where interest is high. McKinsey similarly finds most organizations still experimenting rather than running agents in production.

The honest read for 2026: the current phase is augmentation, not autonomy. AI supports discrete tasks — drafting, summarization, code testing, incident triage — layered onto existing processes rather than redesigning them. Agents are being piloted widely and trusted narrowly. The organizations pulling ahead are those redesigning workflows around AI, not the ones bolting a chatbot onto an unchanged operating model.

Why haven't most companies captured value yet?

The defining tension of 2026 is the gap between adoption and impact. McKinsey's State of AI 2025 found that while 88% of organizations use AI, only about a third have begun to scale it, and just 39% of respondents attribute any EBIT impact to AI — most of them putting that impact below 5% of EBIT. Widespread use has not yet translated into widespread financial return.

This is the scaling gap: pilots are cheap and easy; rewiring an organization is neither. The blockers are consistent across surveys — data readiness, governance and risk controls, workflow redesign, and talent. The winners of the next phase will be defined not by whether they use AI, but by whether they can operationalize it: move it out of the sandbox, into core processes, and onto the income statement.

Where is adoption highest — and why do the Nordics lead Europe?

Within the EU, AI adoption is a Nordic story. Eurostat's 2025 figures put Denmark first at 42.0% of enterprises, followed by Finland at 37.8% and Sweden at 35.0% — all roughly double the EU average of 20% (Eurostat, December 2025). Denmark also posted the single largest year-on-year jump, up 14.5 percentage points.

The Nordic lead reflects long-standing digital infrastructure, high cloud penetration, digitally skilled workforces and strong public-sector digitalization — the same foundations that made the region an early mover in e-commerce and digital government. For Sweden specifically, third place in the EU is a strong position but not the top: Denmark and Finland are adopting faster. The competitive question for Swedish business in 2026 is less whether to adopt AI and more whether it can close the gap with its Nordic neighbors on scaling and value capture.

How many jobs are actually exposed to AI?

Exposure is not the same as replacement, and the leading estimates measure different things. The IMF, in its January 2024 staff discussion note Gen-AI: Artificial Intelligence and the Future of Work, estimates that about 40% of jobs worldwide are exposed to AI, rising to roughly 60% in advanced economies because those labour markets are weighted toward cognitive, task-oriented work. Crucially, the IMF splits that exposure: about half of exposed jobs could be negatively affected, while the other half could benefit from AI-driven productivity gains.

The ILO takes a narrower, task-level view. Its May 2025 study with Poland's NASK institute, Generative AI and Jobs: A Refined Global Index of Occupational Exposure, finds that about one in four jobs (25%) globally is exposed to generative AI, but only around 3.3% of global employment sits in the highest-exposure category. The ILO's headline conclusion is explicit: transformation of job descriptions, not wholesale replacement, is the most likely outcome, because most occupations bundle tasks that still require human input.

Goldman Sachs offers the most-cited automation figure. Its March 2023 analysis estimated that generative AI could expose the equivalent of 300 million full-time jobs to automation globally, while potentially raising annual global GDP by about 7% over a roughly ten-year adoption period. Read together, these three sources agree on direction and disagree on magnitude — a signal that the numbers are estimates of potential, not observed job losses.

Are jobs being created or destroyed on net?

On current employer expectations, created. The World Economic Forum's Future of Jobs Report 2025, based on a survey of over 1,000 employers representing more than 14 million workers across 55 economies, projects 170 million new roles created and 92 million displaced by 2030 — a net gain of about 78 million jobs, equivalent to roughly 22% churn in the labour market.

The composition matters more than the net figure. The WEF identifies technology roles — AI and machine-learning specialists, big-data analysts, fintech engineers — as the fastest-growing, while clerical and administrative roles (data-entry clerks, administrative secretaries, accounting and bookkeeping staff) face the steepest declines. This lines up with the ILO's finding that clerical occupations show the highest generative-AI exposure of any job family.

The net-positive projection carries a large asterisk: 63% of WEF-surveyed employers name skills gaps as the biggest barrier to transformation, and 85% plan to prioritise upskilling. A net gain of 78 million jobs does not mean the same 92 million displaced workers move smoothly into the 170 million new roles — the gross flows imply substantial disruption for individuals even when the aggregate is positive.

Is AI augmenting workers or automating them?

The best real-usage evidence points to augmentation as the dominant mode so far. Anthropic's Economic Index, which analyses millions of anonymised Claude conversations, finds that roughly 57% of measured AI use is augmentation — the model collaborates with, iterates alongside, or advises a human — versus about 43% automation, where the model completes a task with minimal back-and-forth (2025).

That split is context-dependent, and the same data complicates any simple story. Anthropic reports that enterprise API usage skews far more toward automation than consumer chat usage, meaning the augmentation-heavy headline reflects how people use AI today, not a ceiling. As AI is embedded into workflows and agents rather than chat windows, the automation share can rise.

Augmentation and automation are also not mutually exclusive over time. A task that is augmented today — a support agent drafting replies with AI assistance — can become partially automated tomorrow as tools improve and confidence grows. The honest reading of the 2026 evidence is that AI is currently changing how work is done far more than it is removing the worker, but the trajectory is not fixed.

What do the productivity numbers actually show?

The clearest gains are task-level, not economy-wide. The landmark field study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, Generative AI at Work (NBER 2023, published in the Quarterly Journal of Economics 2025), tracked 5,179 customer-support agents and found access to an AI assistant raised productivity — issues resolved per hour — by 14% on average, with a 34% gain for novice and lower-skilled workers and little effect on the most experienced. The mechanism: AI disseminates the tacit know-how of top performers, compressing the experience curve.

Stanford HAI's AI Index 2026 corroborates the pattern across domains, reporting productivity increases in the 14–26% range for customer support and software development, with weaker or even negative effects on tasks that require judgement. It also flags the first concrete labour-market cost: entry-level software-developer roles for workers aged 22–25 have fallen nearly 20% since 2024 — the first white-collar category to show a measurable AI-attributable contraction.

Adoption is now broad enough to matter. Gallup found the share of US employees using AI at work at least a few times a year rose from 40% to 45% between Q2 and Q3 of 2025, and the St. Louis Fed reports US work-adoption of generative AI climbed to 37.4% by August 2025. Wide adoption plus measurable per-task gains has not yet produced a visible jump in aggregate productivity statistics — the classic Solow-paradox lag between deployment and measured output.

What does the evidence not yet show?

It does not show broad, AI-caused unemployment. As of 2026, the strong effects are measured at the level of tasks and specific occupations — customer support, entry-level coding, clerical work — not across the whole labour market. The IMF, ILO and WEF figures are projections and exposure estimates; they are not counts of jobs already lost.

It also does not yet resolve the wage story. The evidence is mixed and early: the same tools that lift novice productivity by 34% (Brynjolfsson et al.) can compress skill premia by making less-experienced workers more competitive, while other analyses warn AI could widen inequality if gains concentrate among capital owners and high earners — the IMF's central concern in 2024. Which effect dominates depends on adoption speed, policy and how firms redistribute productivity gains.

The reskilling imperative is the one point of near-consensus. With the WEF projecting that a large share of core skills will change by 2030 and 85% of employers prioritising upskilling, the binding constraint on a net-positive outcome is transition capacity, not the technology itself. The most defensible 2026 conclusion: AI is reshaping work faster than it is destroying it, but the distribution of who wins and who is left behind remains genuinely open.

Who actually leads the frontier — the US or China?

The United States leads the frontier of AI in 2026, but its lead is narrower than the headline capital figures suggest. On the inputs that produce frontier models, US dominance is clear. American organisations released 59 notable AI models in 2025 against China's 35, according to the Stanford HAI AI Index 2026. The US hosts roughly three-quarters (74.5%) of global AI-supercomputer performance to China's ~15% (14.1%), a ratio of about five to one, per Epoch AI's May 2025 dataset. And US private AI investment reached $285.9bn in 2025 — 23 times China's $12.4bn (AI Index 2026).

On outputs, the picture flips. The performance gap between the best US and Chinese models on standard benchmarks fell to 2.7% by March 2026, down from between 17.5% and 31.6% in early 2023 (AI Index 2026). The inflection point was DeepSeek-R1, which matched the leading US model in February 2025 at a fraction of the reported training cost. The strategic read for 2026: money and compute still concentrate in the US, but capability has commoditised faster than the capital gap would predict. Leading on inputs no longer guarantees a durable lead on the product users actually touch.

Where does China lead outright?

China leads the world on research volume and intellectual property. It accounts for 23.2% of global AI publications and 20.6% of AI research citations, and its share of the 100 most-cited AI papers rose from 33 in 2021 to 41 in 2024 (AI Index 2026, R&D chapter). On patents, China holds roughly 60% of global AI patent grants and filed close to 70% of AI patents worldwide, though the US retains a lead in high-impact, frequently-cited patents. WIPO's 2026 data shows China-based inventors published over 43,000 generative-AI patent families in 2024-2025 combined, more than the country's entire 2014-2023 output.

China's second edge is open weights. Where the top US labs increasingly gate their strongest systems behind APIs, Chinese labs — DeepSeek, Alibaba's Qwen, Moonshot, ByteDance — have made powerful open-weight models the default, seeding global developer adoption and downstream citation. For a citation-driven economy, this matters: open weights travel further into research, tooling and the training data of the next generation of models. China is not winning the frontier, but it is winning distribution and academic footprint — and it is doing so while spending an order of magnitude less capital.

Can Europe compete without frontier labs?

Europe is running a different race — deliberately. It has almost no frontier labs (France's Mistral is the notable exception), and neither the compute nor the private capital to challenge US and Chinese model-builders head-on. What Europe leads is adoption and rules. The AI Index 2026 records Europe, alongside China, posting the highest year-over-year increases in organisational AI adoption, and the EU AI Act — in force since August 2024, with general-purpose-AI obligations applying from August 2025 — makes the bloc the world's regulatory reference point.

This is Europe's 'third way': not building the models, but setting the terms of use and leading in enterprise diffusion. The bet is that governing and applying AI responsibly is itself a durable position, exporting European standards the way GDPR did for data. The risk is structural dependence — running the global economy's most-regulated AI on American and Chinese models. For European operators the practical takeaway is that competitive advantage in 2026 comes from deployment and compliance capability, not from owning the underlying model.

Who else matters — the UK, India, the Gulf and Asia's small giants?

Below the top two, a second tier is consolidating. The United Kingdom keeps genuine frontier-adjacent research strength through Google DeepMind (London) and a well-funded AI Safety Institute. India competes on talent and scale of deployment rather than model-building, and hosts a fast-growing developer base. The Gulf states — the UAE and Saudi Arabia — are buying their way in with sovereign compute and capital: the UAE's Falcon models and vehicles like MGX and Humain pair petro-capital with GPU access.

Asia's smaller economies punish the assumption that adoption follows model-building. On the AI Index 2026's measure of generative-AI adoption reaching the population within three years, Singapore (61%) and the UAE (54%) lead the world, while the US ranks only 24th at 28.3%. South Korea combines strong semiconductor supply (Samsung, SK Hynix) with domestic models from Naver and LG. The lesson for 2026: you do not need to build a frontier model to be an AI power — sovereign compute, talent density and aggressive adoption are separate, winnable games. This is why 60-plus countries now run national AI strategies, tracked across 70+ jurisdictions and 1,000+ policy initiatives by the OECD.AI Policy Observatory.

Where do Sweden and the Nordics fit?

The Nordics are a sharp illustration of the adoption-versus-creation split. Denmark, Sweden and Finland rank among the EU's highest for enterprise AI use in Eurostat's business surveys, and the region's cheap green electricity, dense data-centre footprint and high-trust digital public infrastructure make it a natural place to deploy AI at scale. On the demand side, the Nordics are world-class.

On the supply side, they build essentially zero frontier models. There is no Nordic lab in the AI Index 2026's tally of notable systems; the region's flagship effort, Sweden's GPT-SW3 language model from AI Sweden, is a valuable sovereign-language asset, not a frontier competitor. The Nordic position mirrors Europe's in miniature — lead on responsible, high-productivity adoption; depend on foreign models for the underlying capability. For Swedish operators the strategic question for 2026 is not whether to build a frontier model (they will not) but whether world-leading adoption can be converted into durable economic advantage before the compute and capital gaps harden the global hierarchy.

What does the EU AI Act actually require?

The EU AI Act (Regulation 2024/1689) is the world's first comprehensive, horizontal law on artificial intelligence. It entered into force on 1 August 2024 and applies a risk-tiered model: a handful of practices are banned outright (such as social scoring and untargeted facial-recognition scraping), "high-risk" systems in areas like hiring, credit and medical devices face strict obligations on data quality, documentation and human oversight, and general-purpose AI (GPAI) models carry transparency and, for the most capable models, systemic-risk duties.

The rules phase in over three years. Prohibited-practice and AI-literacy obligations became applicable on 2 February 2025; GPAI obligations and the penalty framework followed on 2 August 2025. The high-risk regime for Annex III systems was originally set for 2 August 2026, but under the Commission's "Digital Omnibus" simplification package that deadline is being deferred toward December 2027 — a signal that even Europe is recalibrating the pace of enforcement.

Penalties are steep by design. Under Article 99, breaches of the prohibited-practice rules can draw fines up to €35 million or 7% of worldwide annual turnover, whichever is higher; other violations reach €15 million or 3%, and supplying incorrect information to authorities up to €7.5 million or 1%. This is not financial advice, but for global firms the Act functions as a de facto standard — much as GDPR did for privacy.

How do the US and China approach AI differently?

The United States has swung toward deregulation. On 23 July 2025 the White House released "Winning the Race: America's AI Action Plan," a package of more than 90 federal actions built on three pillars — accelerate innovation, build AI infrastructure, and lead in international AI diplomacy and security. It arrived with three executive orders, including one fast-tracking data-centre permitting and one requiring federal agencies to procure only models judged "objective" and free of ideological bias. There is no single federal AI statute; oversight is spread across sector regulators and a patchwork of state laws.

China takes a third path: permissive on industrial deployment, strict on content and information control. Its Interim Measures for generative AI services (effective August 2023), deep-synthesis provisions and algorithm-recommendation rules require security assessments, filing of algorithms, labelling of synthetic media and adherence to "core socialist values." The result is three distinct regulatory philosophies — rights-based (EU), market-and-security-led (US), and state-directed (China) — that global developers must reconcile simultaneously.

Are AI harms actually increasing?

Yes, on the available evidence. The AI Incident Database recorded 362 incidents in 2025, up 55% from 233 in 2024, according to the Stanford HAI AI Index 2026. The OECD's AI Incidents and Hazards Monitor, which uses a broader automated pipeline, peaked at 435 incidents in a single month (January 2026). Some of this reflects better reporting rather than pure growth in harm — but the trend line is unambiguously up as deployment reaches 88% enterprise adoption.

The harms cluster around a few themes. Deepfakes and synthetic media drive fraud, non-consensual imagery and election-related manipulation. Bias in high-stakes systems — hiring, lending, policing — remains persistent, and the AI Index notes that responsible-AI benchmarks and safety evaluations are not keeping pace with capability gains. For a business audience the practical exposure is concrete: model errors, data leakage and reputational incidents are now operational risks to be managed, not hypotheticals.

How worried is the public — and can AI be made safe?

Sentiment has cooled markedly. Pew Research Center found that 50% of Americans are more concerned than excited about the increased use of AI in daily life (June 2025), up from 37% in 2021, while just 10% are more excited than concerned. Trust also splits sharply by geography: the 2025 Edelman Trust Barometer reports 72% of people in China trust AI versus only 32% in the US, with older, lower-income and female respondents least trusting. Edelman's central finding is that hands-on experience is the fastest route to trust.

On the safety-engineering side, the field has professionalised. A network of government AI Safety (now often "AI Security") Institutes — led by the UK and US and coordinated internationally — conducts pre-deployment testing, and frontier labs publish safety frameworks (variously called Responsible Scaling Policies, Preparedness or Frontier Safety Frameworks) that tie model release to red-teaming and capability thresholds. Techniques such as alignment training, adversarial red-teaming and evaluations for dangerous capabilities are now standard practice, even as independent auditors warn that disclosure and transparency lag behind the labs' own capability progress.

Near-term versus catastrophic risk: which debate matters?

The safety community is split, and both camps are worth taking seriously. One view prioritises near-term, measurable harms — bias, deepfakes, fraud, labour displacement, privacy and concentration of power — arguing these are happening now and are addressable with existing tools like auditing, disclosure and the EU's risk tiers. The other emphasises catastrophic or existential risk from highly capable, poorly aligned systems, and pushes for frontier-safety frameworks, compute governance and pre-deployment testing before capabilities outrun control.

In practice the two agendas overlap more than the rhetoric suggests: red-teaming, evaluations, transparency requirements and incident tracking serve both. There is also a real environmental cost to weigh. The IEA projects data-centre electricity demand rising from roughly 415 TWh in 2024 (about 1.5% of global consumption) to around 945 TWh by 2030 (just under 3%) — comparable to Japan's current annual use — with AI-driven "accelerated servers" growing about 30% a year. Governance in 2026 is therefore not one debate but several, running in parallel: rights, security, reliability and resources.

Does the scaling still pay off?

Two cost curves are moving in opposite directions, and both matter. On the training side, the compute used for frontier models has grown roughly 4.1x per year since 2010, and Epoch AI estimates the amortised cost of the largest training runs has risen about 2.4x annually since 2016 — putting billion-dollar runs within reach by 2027. On the inference side, the price of a fixed capability is collapsing: Stanford's AI Index found the cost of GPT-3.5-level output fell more than 280-fold between late 2022 and late 2024, helped by roughly 30% annual hardware-cost declines and 40% annual energy-efficiency gains.

The practical implication for 2026-2027: building the newest frontier model gets dramatically more expensive, while using last year's frontier capability gets dramatically cheaper. That asymmetry favours a small number of well-capitalised labs at the leading edge and a much larger field of applications built on rapidly cheapening inference. It also means capability-per-dollar for most real-world use keeps improving even if headline model costs rise — a point often lost in bubble commentary.

Will agentic AI move from demo to deployment?

The clearest product direction for 2026-2027 is agentic AI — systems that plan and execute multi-step tasks with tools rather than just answering prompts. Every major lab roadmap (OpenAI, Anthropic, Google DeepMind) and analyst house has pushed agents as the next commercial wave, and coding, customer support and research are the early beachheads where the economics already work.

The gap between demo and deployment is reliability, and it is the honest constraint here. Multi-step agents compound errors: a task requiring twenty sequential steps at 95% per-step reliability succeeds only about a third of the time. That is why enterprise adoption has lagged the hype — the widely cited MIT report found ~95% of GenAI pilots delivered no measurable P&L impact, with value concentrated in the narrow set of workflows companies actually integrated deeply. Expect 2026-2027 to be less about smarter agents and more about the unglamorous scaffolding — evaluation, guardrails, human oversight, and picking bounded tasks — that turns capability into dependable output.

Is energy the real ceiling?

Increasingly, the binding limit is not model quality or even money — it is power. The IEA projects data-centre electricity demand roughly doubling from ~485 TWh in 2025 to ~950 TWh by 2030, about 3% of global electricity, with AI-focused capacity tripling. In 2025 data-centre demand rose ~17% against ~3% overall power-demand growth. Epoch AI projects frontier training power needs growing 2.2-2.9x per year, with individual training clusters plausibly reaching 4-16 GW by 2030 — the scale of several nuclear reactors.

Grids move slowly. Interconnection queues, transmission build-out and generation additions run on multi-year timelines that do not match AI's compute ambitions, which is why capacity — not chips — is emerging as the near-term bottleneck in the US and parts of Europe. This is the most physically grounded prediction in the outlook: whatever the models can do, in 2026-2027 what actually gets built will be gated by where the megawatts are. Watch power-purchase agreements, nuclear and gas commitments, and grid-connection dates as leading indicators of real AI supply.

Bubble or buildout — can revenue catch the capital?

The defining financial tension is the gap between spending and earning. Analysts estimate the five largest Western hyperscalers will spend roughly $725 billion on capex in 2026, up about 77% from ~$410 billion in 2025, the bulk of it AI infrastructure. Multiple analyses put the annual gap between AI infrastructure spend and AI ecosystem revenue in the hundreds of billions of dollars. That does not by itself prove a bubble — infrastructure precedes revenue in every major build-out — but it does define the risk.

What would have to be true for the buildout to sustain: inference demand keeps compounding, agentic products convert into recurring enterprise revenue, and the depreciating GPU fleet earns back its cost before the next generation obsoletes it. What would have to be true for a bust: revenue stalls while capex is already committed, debt-financed data centres come online into soft demand, and a few anchor customers retrench. The most likely 2026-2027 path is neither euphoria nor collapse but repricing — capital getting more selective, weaker projects cut, and the market rewarding demonstrated revenue over promised capability. As one framing puts it, AI can change the world and still be a bubble; both can be true at once.

What tightens, what splits, and what stays unknown?

Regulation is tightening but unevenly. The EU AI Act's general-purpose model obligations applied from August 2025, and its high-risk rules were originally set for 2 August 2026 — but a May 2026 'Digital Omnibus' provisional agreement postponed the stand-alone high-risk (Annex III) obligations to 2 December 2027 and product-embedded ones to August 2028. So the regime is real and getting stricter, but the hardest compliance dates have slipped; treat any single date as provisional until published in the Official Journal.

Two structural splits will define competition. Open-weight vs closed: open models have closed much of the quality gap (the AI Index measured the top gap narrowing from 8% to 1.7% in a year), pushing closed labs to compete on frontier capability and integration. US vs China: Chinese labs, led by efficient open releases like DeepSeek's, have shown near-frontier results at lower disclosed cost, and export controls plus a bifurcating compute supply chain make this a geopolitical contest, not just a commercial one.

Finally, the honest unknowns. AGI timelines are contested — credible researchers span a decade or more in either direction, and we do not forecast them. Whether reliability improves fast enough for high-stakes autonomy, and whether the economic payoff shows up in aggregate productivity data (which through 2025 remained hard to detect), are open empirical questions. The trajectory is clear; the destination and the date are not.