IV. The Widening Gap

Every paper in the series will include an update. The interval between papers is four months. That is long enough for the AI field to move materially, and the analysis of any domain has to account for where the field is rather than where it stood when the previous paper appeared. The update is not a postscript. It is part of the method. Each paper will revisit the framework's earlier arguments, test them against intervening developments, and state openly where the analysis has held, where it requires revision, and where events have overtaken it.

This is the first such update. It covers the roughly eighteen months during which Code After was developed from a single academic paper, to a fifteen-chapter manuscript, to v0.9, and the period since v0.9's release on 12 April 2026.

The G2 and G3 framing has organised the manuscript's analysis of jurisdictional position from the project's beginning. G2 names the two countries, the U.S. and China, that possess the full stack of frontier AI capacity at scale; G3 names the broader set including the EU, whose regulatory and market-access leverage shapes how AI systems are deployed globally.

It is the framework's most-tested analytical claim. Over the eighteen months during which the project has been in development, the framing has not weakened. It has hardened. The concentration of frontier AI capacity within a small number of G2 labs continues, and the advantages of those labs compound while the distance between them and every other actor widens. All three are more clearly observable now than they were when the project began. The series will therefore continue to use the G2 and G3 framing, because the evidence of the period has provided no reason to abandon it and several reasons to hold it more firmly.

The gap is not only one of capability. It is one of institutional position relative to the frontier — the structural feature Section III describes as decoupling, visible now in four dimensions.

Capital

The scale of funding required to operate at the frontier continues to increase, and at exceptional speed. Valuations and capital requirements for leading foundation-model firms have reached levels that only the U.S. and China can accommodate at scale. The effective unit of analysis is no longer the country; it is the small number of labs within those two countries large enough to be capitalised at all.

The comparison to Europe's most prominent frontier firm makes the scale concrete. Mistral's September 2025 Series C closed at a post-money valuation of approximately €11.7 billion — a serious figure, reflecting both real European ambition and real European capability.

OpenAI's $122 billion funding round in March 2026 valued the firm at approximately $852 billion post-money. Anthropic's $30 billion Series G in February 2026 valued it at approximately $380 billion post-money.^[1]

The gap between Mistral and the leading U.S. labs is roughly 28x relative to Anthropic and 62x relative to OpenAI. It is a gap in capital, not in technical ambition or engineering quality. If OpenAI and Anthropic complete the public offerings widely reported across mainstream financial press for late 2026, that capital gap is likely to widen further.

Capital at this scale concentrates almost entirely in the U.S. and China, mobilised by different institutional routes — primarily private capital markets in the U.S., a combination of state-aligned investment and major commercial actors in China. Actors elsewhere, including private firms in either country operating without those resources behind them, compete under structurally different conditions.

The capital concentration is reinforced by infrastructure. Frontier model firms sit on top of cloud and data-centre buildouts whose scale now rivals major public works. Q1 2026 earnings, reported on 29 April, confirmed the trajectory. The four U.S. hyperscalers — Microsoft, Alphabet, Amazon, Meta — spent on the order of $130 billion on capital expenditure during the quarter, roughly 80 percent above the figure for Q1 2025.

Combined 2026 capex guidance from the four firms now sits in the range of $650 to $700 billion, nearly double the approximately $410 billion they spent across 2025. Around three-quarters of the 2026 figure is directed at AI infrastructure — chips, servers, networking, data centres, and the power required to run them. Combined 2026 capex is on track to approach the level of the four firms' combined operating cash flows, against a long-run average closer to 40 percent.^[2]

These are not marginal adjustments to an established capital cycle. They are the outline of an infrastructure regime that only a handful of U.S. and Chinese actors can enter at scale. This is the decoupling morphology operating at the level of capital. The capacity to build the systems that increasingly govern modern life is concentrated in a number of operators that can be counted on two hands. Authority migrates, in the precise sense the framework names, to the actors who possess that capacity. The institutions formally responsible for governing the resulting systems lack the resources to observe them at the scale at which they are being built, let alone to fund or replicate them.

Compute and Inference

The binding constraint is shifting from training to inference. Earlier phases of frontier development were limited primarily by training compute; current constraints are increasingly defined by the cost and availability of serving models at scale. Usage limits on the consumer products of the frontier labs have tightened over the past year, and plan tiers have been restructured upward, some substantially.

OpenAI has withdrawn Sora's consumer access over the course of 2026, discontinuing the web and app experience in April and the API later in the year. The withdrawal is consistent with the broader inference-capacity pressures visible across frontier deployment, though OpenAI has not publicly framed it in those terms. The feedback loop v0.9 described is now visible in operation. Usage produces revenue and data; revenue and data fund compute; compute enables more capable models; more capable models attract further usage. The frontier is extending its lead through the same mechanism that created it.^[3]

The Application Layer and Agents

Agentic systems have moved from frontier novelties into broader consumer and enterprise integration. The pattern is visible at two levels.

At the consumer level, Manus was among the first widely visible general-agent products from the Chinese ecosystem. In 2025 it drew intense attention in China — domestic coverage framed it as a "GPT moment for agents" and as the second major Chinese AI breakthrough of the year after DeepSeek — establishing it as a focal example of how rapidly agent capability had crossed into ordinary consumer use.

The case then took on a second, geopolitical, dimension. Meta announced its acquisition of Manus in late December 2025 in a deal reported at approximately $2 billion. On April 27, 2026, China's National Development and Reform Commission (NDRC) announced that it would prohibit foreign investment in Manus, citing foreign-investment and national-security grounds, and ordered the parties to unwind the transaction.

The case carries a dual lesson. The first half is the one already visible: agent capability is reaching ordinary users at speed. The second is newer. Sovereign action is now drawing, in real time, the cross-border boundaries of access to frontier AI capability. Within the broader process of U.S.–China decoupling in AI, the Manus case marks a clear inflection point: the cross-border flow of frontier AI assets is no longer determined by market logic alone, but is increasingly subject to the hard-edged constraints of national-security review.

At the infrastructure level, the pattern presents differently. Open-source agentic frameworks have emerged as a distinct category. OpenClaw is the most visible example. Uptake within China's developer community and beyond has been rapid. Reports describe nearly a thousand people queuing outside Tencent's Shenzhen headquarters for installations. Local governments have subsidised deployment events. Cloud providers and regional actors have forked and repackaged the codebase at scale. Audited user metrics remain limited, but the pace of replication and the intensity of public and developer discourse around the framework suggest that market reception has far exceeded initial industry expectations.

OpenClaw functions as the Orchestration Layer between foundation models and user-facing applications, managing state and memory, tool and API access, permissions, and routing across multiple model backends. It does not compete with frontier models. It makes them work. The orchestration layer is the path by which the concentrated capability sitting at the top of the stack reaches the users, applications, and workflows at the bottom.^[4]

The emergence of this middle layer brings the real shape of decoupling into focus. The AI stack now resolves into three distinguishable layers. The foundation layer, where a small number of frontier models sit, is consolidating. The orchestration layer, where frameworks like OpenClaw assemble models into systems of action, is becoming more open and portable. The application layer, where users meet these systems through products, interfaces, and workflows, is diffusing at speed.

This is no coincidence. The opening of the orchestration layer and the diffusion of the application layer do not disperse frontier capability; they deepen dependence on it. Frameworks like OpenClaw, by routing agents across multiple providers, reduce vendor lock-in at the orchestration level — but the quality of system cognition still depends on the quality of available models. Orchestration can amplify, structure, and stabilise intelligence. It cannot manufacture frontier-grade intelligence from non-frontier foundations. The easier it becomes to build agentic systems on top of the best frontier models, the stronger the incentive to use those models rather than to build alternatives. Application diversity has not distributed AI sovereignty. It has widened the demand funnel through which centralised intelligence is further reinforced.

The shift from conversation to action changes the stakes. Early AI systems answered questions and humans did the work; agentic systems increasingly do the work themselves — drafting and sending emails, scheduling and rescheduling meetings, organising files, moving funds, completing transactions. The user sees the result. The user does not see the chain of decisions and tool calls that produced it. Once AI not only recommends action but takes it, the distance between what the user sees and what the system does becomes a question this series will return to repeatedly in subsequent papers.

The application layer is expanding. The orchestration layer is opening. The foundation layer is consolidating. These three motions push the concentration v0.9 described in the same direction — deeper, not wider.

World Models and Physical AI

The same pattern of concentration is extending beyond language. Work on systems that model and predict the physical world is moving from research into early product discussion, on three parallel tracks. Major frontier labs including xAI, Meta, and Google DeepMind are extending their language-model capability into the physical domain, with humanoid programmes such as xAI's Optimus push moving from prototype toward deployment. The Chinese ecosystem is doing parallel work, anchored by Unitree, the broader humanoid manufacturing base, and industrial-AI deployment programmes already operating at scale.

A third track has emerged through dedicated start-ups built specifically around world-model construction, establishing themselves as a distinct strategic line in the field. Two of the most credentialed researchers in artificial intelligence have stepped away from their institutions to lead this work. Fei-Fei Li — whose ImageNet project a decade earlier catalysed the deep-learning era — has taken extended leave from Stanford to found World Labs, a spatial intelligence company whose core mission is the construction of large-scale world models.

Yann LeCun, recipient of the A.M. Turing Award for his foundational work on deep learning, has departed Meta and co-founded AMI Labs (Advanced Machine Intelligence) to pursue the same class of systems. His public position is that large language models alone cannot reach artificial general intelligence. AGI is the threshold at which a system exhibits cross-domain autonomous reasoning and generalisation comparable to human cognition. World models, LeCun argues, are the architecture that can.

This view is no longer confined to research-side advocacy. In April 2026, the Goldman Sachs Global Institute framed world models as a decisive next step in AI and argued that current AI-infrastructure projections, built around language-model scaling, materially understate the capital, simulation, and physics-engine demands of the world-model track. In the same week, MIT Technology Review's new "10 Things That Matter in AI Right Now" list identified world models as one of the field's most consequential current developments. The thesis that the next layer of AI capability lies beyond language is now converging across research labs, finance, and technology press — within weeks of this paper's planned publication.

World models train on different data altogether. They learn from what cameras see, what microphones hear, what touch and motion sensors record across thousands of synchronised channels — visual, acoustic, kinetic, and physical streams drawn from real and simulated environments rather than from text. The sensing, modelling, validation, and deployment infrastructure these systems need exists today in fragments. It is not yet at the scale or maturity that broad deployment will require.^[5]

Humanoid robotics makes the point concrete. Over the past eighteen months, Unitree, Figure, Tesla, and the broader humanoid ecosystem have moved these systems from laboratory curiosity to commercial roadmap. Chinese activity is especially visible — factory pilots, training centres, industrial-policy initiatives. Humanoid robots are not world models. They are among the most important physical platforms through which the multimodal embodied data world models need can be generated at scale, alongside simulation environments, industrial sensor networks, and autonomous vehicles.^[6]

Language-model capacity already concentrates in the G2. World-model capacity is likely to concentrate further. The data, the robotics manufacturing base, and the sensor infrastructure all concentrate there more strongly still. The dynamic that shaped the language-model era is extending into the era of physical intelligence. The gap v0.9 described is not stabilising at the language-model frontier. It is moving.

This is what makes the series' first paper urgent rather than thematic. The Language paper is not a neutral topic. Language is where the gap first opened. Language is where everything else rests. A country that cannot govern in its own language against AI systems that think in another will not govern the embodied, physical-world systems that come next. The same is true for every domain the series addresses. The work begins with language because language is where the gap first became visible and where it is still most correctable.

The framework has held. A reviewer could reasonably have argued a year ago that the G2 framing overstated a temporary condition. The evidence since does not support that reading. The gap is widening along every dimension the framework identified — capital, compute, data, talent, models, infrastructure — and now extending into a further dimension, physical-world data, that will widen it more.

The decoupling described in Section III is intensifying, not normalising.

Each subsequent paper will open with an update of this kind. The framework will be tested against the previous four months. The analysis will state openly where it has held and where it has had to evolve. AI does not wait. The project will not pretend it does.

The updates will track both G2 ecosystems on equal terms, the U.S. and China, without filtering one through the lens of the other. Most current AI analysis treats one ecosystem as the default frontier and the other as competitive context. Code After treats both as the frontier, because both are.

The project exists to connect what is happening at the frontier to the lives AI is rebuilding. It is happening to everyone at once — the lawyer, the accountant, the doctor, the teacher, the CEO, the politician, and the citizen scrolling the news on their phone. The reach is the goal. The series is built to deliver the analysis to readers who need it, in the languages they read in, at a pace that matches what is actually changing around them. How the series is built to do this is the subject of the sections that follow.

Code After is a framework, a series, and a living record of its own claims against what actually happens.

1. Mistral AI, Series C announcement (September 2025); OpenAI, funding round announcement (March 2026); Anthropic, Series G announcement (February 2026). Valuations as reported by primary company communications and corroborated in standard technology-press coverage. ↩︎
2. Capital expenditure figures drawn from the Q1 2026 earnings releases of Microsoft, Alphabet, Amazon, and Meta (29 April 2026) and the corresponding SEC Form 8-K filings, with mainstream financial-press aggregation used for full-year guidance synthesis. Q1 2026 combined capex on the order of $130 billion, against approximately $72 billion in Q1 2025, derives from per-company filings (cash payments for property, plant, and equipment, with finance leases included where reported). The 2026 full-year guidance range of roughly $650 to $700 billion synthesises updated investor guidance issued the same week. The 2025 baseline of approximately $410 billion is consistent with mainstream financial-press aggregation. The AI-infrastructure share (approximately three-quarters of 2026 spend) and the capex-to-operating-cash-flow ratio (approaching 100 percent in 2026, against a long-run average closer to 40 percent) reflect standard analyst aggregations, including UBS, Bank of America, Epoch AI, and other major sell-side and independent research houses. ↩︎
3. OpenAI, Sora consumer access withdrawal (April 2026); API discontinuation announced for September 2026. Company communications and standard technology-press coverage at the time of release. ↩︎
4. On Manus's March 2025 launch and immediate public reception in China, see China Daily (6 March 2025) describing the product as "the GPT moment for AI agents" and as the second major Chinese AI breakthrough after DeepSeek, and Tech Xplore (March 2025) on the invitation-only debut and ensuing public reaction. On Meta's late-2025 acquisition announcement, the subsequent regulatory review by Chinese authorities, and the NDRC's formal prohibition of the transaction, see Bloomberg (27 April 2026) on the NDRC's decision to prohibit foreign investment in Manus and to order cancellation; CNBC (27 April 2026) on the NDRC's role as state planner and on the framing of the action under foreign-investment and national-security grounds; and parallel reporting across major Western financial and technology press, including Wall Street Journal, The Washington Post, Reuters, and TechCrunch, on the same date, alongside coverage in Chinese state and financial media including Caixin and China Daily and the NDRC's own statement of the same date. On OpenClaw, see Chinese and international technology and financial media coverage of its rapid 2026 adoption — including reporting on queues for installations at major Chinese technology firms, local-government subsidies for deployment events, and investor enthusiasm — together with documented forking and repackaging of the codebase and its precursors (notably Clawdbot and Moltbot lineages) on GitHub. Adoption here refers to observable ecosystem replication — forks, derivative projects, developer commentary, media coverage — rather than to audited user metrics, which remain limited. ↩︎
5. On Fei-Fei Li's foundational work, see J. Deng et al., ImageNet (2009), and A. Krizhevsky, I. Sutskever, and G. Hinton, AlexNet (2012). On Li's 2024 leave from Stanford and the founding of World Labs, see World Labs' launch communications and September 2024 coverage of its USD 230 million funding round in Forbes, Reuters, and TechCrunch. On Yann LeCun, see ACM, 2018 A.M. Turing Award announcement (March 2019), jointly to Y. Bengio, G. Hinton, and Y. LeCun. On LeCun's late-2025 departure from Meta and the co-founding of AMI Labs, see TechCrunch (December 2025; March 2026), MIT Technology Review (January 2026), and AMI Labs' public materials at amilabs.xyz. On LeCun's published position on LLMs and AGI, see his MIT Technology Review interview (January 2026) and the AMI Labs mission statement. On the convergence of finance-side and technology-press analysis on the world-model thesis, see G. Lee and D. Keyserling, When AI Learns How the World Works (Goldman Sachs Global Institute, April 2026), and MIT Technology Review, "10 Things That Matter in AI Right Now" (April 2026). ↩︎
6. On the frontier-lab physical-AI push, see Tesla and xAI's public communications on the Optimus humanoid robot and the Digital Optimus / Macrohard initiative, Meta's world-model research portfolio (including V-JEPA 2), and Google DeepMind's Gemini Robotics publications. On humanoid commercialisation 2024–2026, see Unitree's product and shipment communications, Tesla's Optimus programme updates, Figure's pilot-deployment announcements, and standard technology-press coverage. On Chinese humanoid deployment, training infrastructure, and industrial-policy support, see Chinese and Hong Kong technology-media coverage in 36Kr, TechNode, and the South China Morning Post, alongside Chinese state-media coverage of national humanoid-industry initiatives. ↩︎