1. ANTHROPIC'S $965 BILLION: THE FINAL NUMBERS AND WHAT THE CAP TABLE SAYS
Anthropic's Series H closed Wednesday at a $965 billion post-money valuation on $65 billion in new capital — the largest single venture financing in the history of the technology industry by a margin that makes the previous record, OpenAI's $122 billion round in March, look like a Series B. The round was co-led by Sequoia Capital, Altimeter Capital, Dragoneer Investment Group, and Greenoaks Capital Partners, with Capital Group, Coatue, D1 Capital Partners, and DST Global among the major participants. Baillie Gifford, Blackstone, Brookfield, D.E. Shaw Ventures, and Fidelity Management & Research came in alongside strategic participants including Samsung, SK Hynix, and Micron — the hardware companies whose products sit at the physical layer of every model Anthropic trains. The presence of the semiconductor manufacturers is not incidental. It is a structural commitment from the companies that make the chips that Claude runs on, timed to arrive at the moment Anthropic has committed to building out up to five gigawatts of AWS Trainium capacity — more than double OpenAI's approximately two gigawatts — and the round's strategic composition reflects that both parties understand the infrastructure build-out and the valuation are the same bet expressed in different instruments.
The revenue basis for the $965 billion figure is specific enough to evaluate. Annualised run-rate revenue crossed $47 billion earlier this month, driven by Q2 2026 revenue projected at $10.9 billion — up 130% from $4.8 billion in Q1. The company reported its first quarterly operating profit of approximately $559 million in Q1, and now counts more than 1,000 enterprise customers spending over $1 million annually. The progression from $9 billion annualised at the end of 2025 to $47 billion in May 2026 is not a normal growth curve; it is a curve that enterprise software businesses observe once in a generation, if ever, and it is almost entirely driven by the shift from experimental AI deployment to production infrastructure at firms like KPMG, PwC, and the Pentagon, where Claude is not supplementing existing workflows but replacing the underlying infrastructure on which professional work runs. The $965 billion valuation implies a forward revenue multiple in line with the highest-valued enterprise software companies in history — companies that, at the time of their comparable valuations, were also either first to a market or had demonstrated the kind of network effect or switching cost that makes their revenue defensible over time. Whether Anthropic's enterprise penetration represents that level of durability is the question the valuation is betting it does.
The investor composition also carries a geopolitical signal that deserves attention separately from the financial analysis. The $15 billion tranche from hyperscalers — principally Amazon's $5 billion commitment from April — represents a portion of the round in which the capital and the strategic alignment are explicitly linked: Amazon is not just a financial investor in Anthropic but the cloud provider on which a substantial share of Claude's inference runs. The participation of Samsung, SK Hynix, and Micron places the semiconductor supply chain inside the cap table of the company that is now the most valuable private AI company on earth. That is a different arrangement from a typical venture round: it is a vertical integration of the AI stack expressed through ownership stakes rather than through acquisition, and it creates a set of aligned interests between model provider, infrastructure provider, and chip manufacturer that is difficult to replicate and difficult to unwind. Whether that alignment is a moat or a constraint — whether having Samsung and Amazon at the cap table helps or limits the optionality of a future IPO — is the structural question the next twelve months of Anthropic's trajectory will begin to answer.
2. KARPATHY JOINS ANTHROPIC: WHAT THE PRETRAINING SIGNAL MEANS
Andrej Karpathy announced Monday that he was joining Anthropic's pretraining team, working under team lead Nick Joseph with a specific mandate to build a new group focused on using Claude itself to accelerate the pretraining research loop. The announcement was made with minimal fanfare — a social media post, a brief public statement — and the understatement is itself informative. Karpathy is not a name that requires a press release to carry weight in the field: he co-founded OpenAI, led Tesla's Autopilot programme through its most technically consequential years, and has spent the intervening period building an audience of hundreds of thousands of engineers through his lecture series on neural network fundamentals. His stated reason for the move — that "the next few years at the frontier of LLMs will be especially formative" and that he wanted to be at Anthropic specifically to do R&D — is a precise enough statement to read as a professional assessment rather than recruiting copy. Karpathy is not joining Anthropic for the salary or the brand recognition; he is joining because he has made a judgment that the next frontier capability advance is going to come from the pretraining work, and that Anthropic's is the team most likely to find it.
The specific structure of his mandate — using Claude to accelerate pretraining research — is a direction that multiple frontier labs have been exploring under different names and with varying degrees of public disclosure, but it has not produced a public breakthrough result yet. The idea is straightforward in concept and extremely difficult in execution: use the capabilities of a current-generation model to automate or assist the research processes that produce the next generation. That includes hypothesis generation, experiment design, code implementation of training modifications, analysis of training dynamics, and identification of scaling regime transitions — all tasks that currently require large numbers of senior researchers working over extended periods. If a frontier model can do a meaningful fraction of that work reliably, the effective research throughput of a small team increases by an amount that is difficult to express in normal productivity terms. The reason it is difficult is that the research tasks in question are exactly the ones where model reliability and honesty matter most: a model that confidently generates plausible-sounding but incorrect hypotheses about training dynamics is worse than no model at all, because it consumes researcher attention on dead ends faster than it generates genuine leads. Karpathy's hire is a statement that Anthropic believes its current models are reliable enough that the risk of that failure mode is manageable, which is itself a signal about where Claude's capability frontier actually sits relative to the tasks it is being asked to perform.
The competitive context of the hire matters beyond its technical implications. OpenAI lost a co-founder to its nearest competitor in a week when that competitor closed a round that overtook OpenAI's private valuation. Karpathy's departure is not a crisis for OpenAI — the company has retained most of its senior research leadership through the IPO preparation period — but it is a data point in the ongoing narrative about which organisation the field's most respected technical minds regard as the better place to do the most important work. That narrative matters for recruiting the next generation of pretraining researchers, for whom the choice of employer is also a judgment about where the field is actually moving. Anthropic has now hired or attracted researchers from OpenAI, DeepMind, and Google Brain in sufficient number that the pattern cannot be attributed to individual decisions. It represents a structural shift in the perceived centre of gravity of frontier AI research that has implications for capability trajectories over the next two to three years that are more consequential than any single benchmark result.
3. CHINA LOCKS ITS AI RESEARCHERS IN: THE TALENT BARRIER AND WHAT IT FORECLOSES
Bloomberg reported Monday that Chinese government agencies have begun imposing formal overseas travel restrictions on private-sector AI researchers at firms including Alibaba, DeepSeek, and other companies working on frontier models. The policy represents a formalisation of soft guidance that had been in place since late 2025: where the earlier approach involved informal advice to prominent AI figures to exercise caution about international travel, the current mechanism requires researchers assessed as strategically significant to obtain approval from "relevant authorities" before leaving the country. The assessment of who qualifies as strategically significant is not based on employment category or seniority level alone; it is based on an individual evaluation of the knowledge and capabilities each researcher holds, which means the policy's reach is determined by an opaque administrative judgment rather than by a clear rule that researchers can orient around. China has applied comparable restrictions to nuclear scientists, aerospace engineers, and certain categories of state-sector executives for decades. Extending the framework to private-sector AI researchers is a significant departure from the implicit operating agreement that has governed the country's commercial technology sector since the era of Alibaba's and Tencent's founding: the agreement that private-sector technologists were not state assets in the way that defence scientists were, and that their mobility was therefore governed by labour law rather than national security considerations.
The strategic logic is legible and has been stated plainly in adjacent contexts by Chinese officials. The concern is not primarily that Chinese AI researchers will defect to Western labs — the historical rate of researcher migration in either direction has been modest — but that international travel, conference attendance, informal collaboration, and the diffuse knowledge-sharing that happens in professional networks will reduce the information asymmetry that Chinese labs currently hold about their own research trajectories. DeepSeek's R1 release in January 2025 demonstrated that the gap between Chinese and American frontier AI development had closed substantially from the position most US observers had assumed. Beijing's response to that demonstration — the travel restrictions among other measures — suggests that Chinese officials regard the information asymmetry as a strategic asset that needs to be actively protected rather than allowed to erode through normal scientific exchange. The travel restrictions are, in this reading, not about keeping researchers in; they are about keeping the knowledge of where Chinese AI development is actually headed from becoming legible to competitors who would use that knowledge to recalibrate their own trajectories.
The consequences for the international AI research ecosystem are structural and largely irreversible once the policy becomes embedded. The field has operated for its entire history on an assumption of relative openness: that researchers in Beijing, London, Montreal, and San Francisco read the same papers, attend the same conferences, and that this circulation of ideas is both a scientific good and a check on the dynamics of arms-race competition. That assumption has been eroding for several years through increasing classification of dual-use research, export controls on chips and model weights, and the gradual bifurcation of research infrastructure along national lines. China's travel restrictions on private AI researchers are the clearest single formalisation of a process that has been under way more diffusely, and they accelerate the trajectory toward a world in which the two most capable AI development ecosystems are effectively opaque to each other at the human level. The geopolitical risk of that opacity is not symmetric: both sides lose the information that would allow them to calibrate whether the other is approaching capability thresholds that warrant a response, which is precisely the condition under which miscalculation becomes most likely.
4. GOOGLE RETIRES THE SEARCH BOX: WHAT HAPPENED ON TUESDAY AND WHY IT IS PERMANENT
On Tuesday May 26, Google completed the transition it had been signalling at I/O the previous week: the company flipped the switch on Gemini 3.5 Flash as the default engine powering core global search, effectively retiring the traditional ten-blue-links interface as the primary way billions of users interact with the web. The change is not absolute — users can still navigate to the link-based results view — but the default experience, the one that most users in most contexts will encounter, is now a conversational interface that synthesises information rather than indexing documents. The traditional search box still exists; what no longer exists is the assumption that typing a query into Google produces a list of links to websites as its primary output. That assumption has been the operating basis of the web's information economy for twenty-five years, and its retirement is not a product update. It is a structural change to how the internet's most important distribution channel works, with consequences that will take years to fully propagate through the web publishing ecosystem, the advertising industry, and the epistemology of how information is attributed, verified, and updated.
The technical choice to run the default search experience on Gemini 3.5 Flash rather than Gemini Omni — the more capable model announced at I/O — is an infrastructure and economics decision that reveals something important about the unit economics of running AI search at Google's scale. Flash delivers intelligence "that rivals large flagship models at speeds expected from the Flash series," in Google's description, but the key phrase is not the intelligence claim; it is the speed and cost profile. Google processes approximately 8.5 billion searches per day. Running each of those searches through a model with the compute profile of Gemini Omni would be financially unworkable at any advertising revenue level; running it through a model optimised for latency and cost at scale is what makes the transition from experimental product to global default economically coherent. The shift to Gemini 3.5 Flash as the search backbone is therefore not a capability downgrade relative to the experimental AI Search that Google has been running in parallel — it is the moment when the capability-cost curve of the Flash-class models crossed the threshold where search-quality intelligence is deliverable at search-scale economics.
For the web publishing ecosystem, the change raises questions that do not yet have clear answers. If Google's default interface synthesises rather than indexes, the traffic that publishers have historically received from users clicking through to source documents is structurally at risk — not because Google blocks access to sources, but because the default user experience no longer requires a click to get an answer. Google's position, articulated repeatedly through the AI Overviews rollout, is that AI-powered search generates more total queries per session and therefore more total clicks to publishers than the traditional model, because users who get partial answers are prompted to explore further. That claim has not been validated at the scale that Tuesday's transition represents, and the web publishing industry — which has already absorbed substantial traffic losses through the AI Overviews transition — is watching the data with an anxiety that has not been fully relieved by Google's assurances. The honest answer is that the evidence is ambiguous and that the new equilibrium, whatever it is, will not look like the old one.
5. OPENAI TURNS ON ADVERTISING: THE MONETISATION SHIFT AND ITS COMPLICATIONS
OpenAI launched a self-serve advertising platform inside ChatGPT this week, allowing brands to pay directly to have their products surfaced within AI-generated responses. The launch follows the confidential S-1 filing from May 22 and represents the first time in ChatGPT's history that commercial promotion has been explicitly and transparently integrated into the product's output rather than monetised exclusively through subscription fees and API access. The platform is self-serve in the sense that brands can configure campaigns through an interface similar to Google Ads or Meta's Advantage+: they specify target queries, creative assets, budget, and the contexts in which their products should appear, and the system handles placement. The key design choice — which distinguishes this from simple banner advertising — is that the placement is inside AI-generated responses rather than adjacent to them. The advertising is meant to appear as relevant information within an answer, not as a clearly delineated ad unit separate from the content the user asked for.
The disclosure and labelling architecture of the platform is the design question on which its legitimacy rests. If users can reliably distinguish between an AI-generated answer that reflects the model's best synthesis of available information and an AI-generated answer that has been influenced by advertising spend, the system is operating with sufficient transparency to be evaluated on its merits. If the distinction is not clear — if a sponsored placement in a ChatGPT response about, say, which project management software to use looks identical to an organic recommendation — then the platform is doing something qualitatively different from search advertising and qualitatively different from banner advertising: it is using the epistemic authority that users assign to AI-generated answers to deliver commercial messages that carry a degree of implied objectivity they have not earned. OpenAI's stated position is that sponsored content will be clearly labelled, but the details of what "clearly labelled" means in practice — how prominent, how persistent, how distinguishable from the surrounding generated text — are not yet public, and they matter more than the stated policy.
The timing relative to the IPO filing is not coincidental, and it should not be treated as such. OpenAI's S-1 will need to show public markets a path to profitability on a revenue base of $25 billion growing at 60% annually but with $14 billion in annual operating losses. Advertising revenue, if the platform achieves meaningful scale, adds a revenue stream with higher gross margins than API inference and lower capital intensity than consumer subscription growth. The question the S-1 will need to answer is whether ChatGPT's 700 million monthly active users constitute an advertising audience comparable in value to Google Search or Meta's feed — whether users interacting with an AI assistant in the context of specific tasks carry the same purchase-intent signal that makes search and social advertising valuable. OpenAI is betting that the answer is yes, and that the intent signal in a ChatGPT query is at least as strong as a Google search query because users are more specific in their AI interactions than in their search queries. That is a plausible hypothesis. It is also one that advertisers have not yet had the opportunity to validate at scale, and the IPO roadshow will be, among other things, an attempt to sell institutional investors on the hypothesis before the data exists to confirm it.