The inference layer raised $763M the week Anthropic raised $65B. One story got told.

Two inference infrastructure deals closed quietly in the same week Anthropic became the most valuable startup on earth. Groq is raising $650M to rebuild as an inference cloud after Nvidia paid $20B for its IP and took its founding team. OpenRouter raised $113M Series B led by Alphabet's CapitalG after five-x-ing weekly token volume in six months, from 5 trillion to 25 trillion. Together: $763M flowing into the layer that delivers AI rather than trains it. Anthropic's $65B round dominated every headline. It should not have obscured the more structurally interesting observation: this was the week the inference delivery layer raised nearly as much as most frontier model labs, and it happened because model proliferation has made routing a real business.

Strip Anthropic out and this week's total is $5.8B across 42 rounds. The prior week (May 18-24) was $5.3B without Anthropic's pre-announcement. Non-Anthropic capital deployed week-over-week: up roughly 9.5%. The underlying flow is steady; it's the megastructure rounds that generate the appearance of volatility. Cognition raised $1B at a $26B valuation. MaintainX was acquired by Autodesk for $3.6B. StackAI was acquired by Asana for $75M. Focused Energy closed the largest fusion Series A in history at $240M. And synthetic data terminology in new startup descriptions jumped 10x in a single week. That's five distinct stories, none of which needed Anthropic to be interesting.

The Numbers

Anthropic's $65B close at a $965B post-money valuation was announced May 28 via Bloomberg and confirmed on Anthropic's own news page. The round is Series H, led by Altimeter, Dragoneer, Greenoaks, and Sequoia, with co-leads including Capital Group, Coatue, D1, GIC, and ICONIQ. It includes $15B of previously committed hyperscaler capital, $5B of which is from Amazon. Annual revenue run-rate has crossed $47B. Anthropic now exceeds OpenAI's last reported valuation, becoming the most valuable private AI company in the world.

The Inference Layer Raised $763M While Nobody Was Watching

The story at Groq is stranger than a simple funding round. In December 2025, Nvidia struck what journalists have called a "not-acqui-hire": a $20B deal that paid out Groq's investors in cash, took CEO Jonathan Ross and president Sunny Madra, and licensed Groq's core LPU hardware technology. Nvidia did not formally acquire Groq as a company. What remained was the inference cloud business, GroqCloud, and a new management team led by CFO-turned-CEO Simon Edwards. Groq is now raising $650M from existing investors Disruptive and Infinitum, who have agreed to backstop the round if other pro-rata holders decline. The FTC announced in January 2026 that it is examining "merger in disguise" arrangements across the tech industry; the Groq deal is a named example of the pattern under scrutiny.

What Groq is raising money to build now is a cloud API that lets developers run inference-heavy workloads across its remaining accelerator infrastructure. The company processed significant inference volume before the Nvidia deal and retains customer relationships. Whether the inference cloud thesis can work without the hardware IP moat that made Groq distinctive is the core question. The $650M says existing investors think it can, and that the customer base justifies the rebuild even without the founding engineering team.

OpenRouter's round landed the same week and represents a different inference bet entirely. Rather than owning compute, OpenRouter routes requests intelligently across 400+ models from different providers. Weekly token volume went from 5 trillion to 25 trillion in six months. The $113M Series B was led by CapitalG (Alphabet's independent growth fund) with participation from NVentures (NVIDIA's venture arm), ServiceNow Ventures, MongoDB Ventures, Snowflake Ventures, and Databricks Ventures. The corporate participation list reads like the enterprise software sector endorsing model routing as infrastructure. When Alphabet, NVIDIA, and four major data platforms co-invest in a router, that is not a bet on one model winning; it is a bet on model proliferation continuing indefinitely.

Groq and OpenRouter are not in the same business, but they share an underlying thesis: the marginal value in AI delivery is shifting from training compute to inference orchestration. One is rebuilding hardware-as-a-cloud after losing the hardware. The other is building the abstraction layer above the hardware. Together $763M flowed into this layer in a single week, mostly unreported while the Anthropic round consumed coverage.

Cognition's Self-Eating Loop at $26 Billion

Cognition closed a $1B Series D on May 27 at a $26B post-money valuation, more than doubling the $10.2B valuation from September 2025. Investors Lux Capital, General Catalyst, and 8VC co-led; Ribbit, Atreides, and Founders Fund participated. The headline number is large. The operational detail is more interesting: Cognition says Devin, its AI coding agent, now writes 89% of Cognition's own internal codebase. Annualized recurring revenue has grown from $73M (disclosed at the September raise) to $492M today, a roughly 6x increase in eight months. Enterprise usage of the product increased tenfold since January.

The "Devin writes its own code" claim has obvious PR value, but it also has a structural implication. If the product genuinely handles most engineering tasks, Cognition's headcount-to-revenue ratio should look anomalous versus traditional software companies. The company has not disclosed employee count, but the pitch is clear: this is evidence that the coding agent category has moved from demo to production at meaningful scale. The distinction between Devin (a cloud agent that handles full tasks end-to-end) and IDE tools like Cursor or GitHub Copilot (which augment developers rather than replace tasks) is increasingly what justifies Cognition's valuation gap. At $492M ARR and $26B valuation, the multiple is approximately 52x forward ARR, which is aggressive but defensible if the autonomous model is producing at the claimed rate.

Enterprise Platforms Are Acquiring Their Way Into the Agentic Era

Two acquisitions announced within 24 hours of each other tell the same story from different angles. Autodesk announced on May 28 that it will acquire MaintainX for $3.6B all-cash, the largest acquisition in Autodesk's history. MaintainX sells maintenance and operations software to factory and facility teams: work orders, inspection records, asset data, day-to-day maintenance logs. The company projects more than $135M in 2026 ARR growing above 50%. Autodesk is buying AI-native maintenance workflows to bolt into its broader industrial lifecycle platform, connecting factory floor operations data to the design and simulation tools Autodesk already dominates.

Asana's acquisition of StackAI for $75M is smaller but structurally similar. StackAI is a no-code platform for designing and deploying AI agents across enterprise systems including Salesforce, Slack, and Google Workspace. Asana had built agents that could plan and track work but could not execute tasks inside external systems. StackAI fills the cross-system execution gap. Asana's public framing described the goal as building a "human-agent OS," the first time a major enterprise platform has publicly committed to that vocabulary. That phrase is load-bearing: it signals that Asana's product roadmap is explicitly about human-agent handoffs rather than pure task automation.

The pattern across both deals: legacy enterprise platforms are buying AI-native point solutions rather than building them internally. At a time when coding agents command 52x ARR multiples, the build-versus-buy math strongly favors buying. This acquisition pattern will accelerate.

Focused Energy Closes the Largest Fusion Series A in History

Not everything this week was software. Focused Energy, a German-American laser fusion company spun out of Technical University Darmstadt, closed $240M in Series A funding on May 27, the largest fully-secured Series A in the global fusion industry to date. Investors include RWE (a potential energy off-taker), Germany's federal innovation agency SPRIND, the European Innovation Council Fund, and returning lead Prime Movers Lab. The capital will be deployed at the former RWE power plant site in Biblis, Hesse.

The investment case for laser fusion improved materially after the NIF demonstrated net energy gain in December 2022, but venture capital at scale has been slow to follow. RWE's participation as both investor and potential buyer of fusion energy output suggests that European energy companies are actively hedging their baseload generation future. AI data center buildout is creating electricity demand that the existing grid is not equipped to meet at projected scale; that urgency is part of what makes fusion investment timelines feel more pressing now than three years ago. The round was oversubscribed and attracted investors from Germany, Europe, Asia, and the Gulf region.

Microtrends Worth Watching

• Synthetic data terminology spiked 10x. Six of the 449 new startups added this week mentioned synthetic data as a core product concept, compared with a normalized rate of roughly one per six days in the trailing four-week period. Foundation model providers restricting training data terms of service is pushing companies needing domain-specific fine-tuning toward generated datasets. The category is moving from technique to startup noun.
• AI red-teaming is now a funded category. Gray Swan Software raised $40M this week. The company runs an arena connecting 15,000 security researchers in competitive red-teaming exercises against frontier models. Its competitions have generated over 250,000 attack attempts across 13 models. This is AI safety not as regulatory positioning but as a recurring enterprise software contract sold directly to the labs.
• Governance layer for agents. Geordie AI raised $30M Series A for security and governance tooling for agentic AI deployments. It is at least the fourth "agent governance" startup to raise a Series A in 2026. Every prior wave of enterprise software adoption generated a secondary governance market roughly 24-36 months after the underlying adoption began. Agentic deployment accelerated in late 2024; the governance market is arriving on schedule.
• Voice agents, quietly multiplying. Three of the 449 new startups this week described voice agents as core to their product, a 3.4x frequency increase versus the trailing four-week rate. The numbers remain small; the direction is consistent with enterprise interest in voice-based automation for customer service and field operations.
• Memory as AI's bottleneck. XCENA's $135M Series B at a $570M valuation is a bet that memory bandwidth, not raw compute, is the binding constraint for large-context AI inference. The company is building semiconductor architecture targeting the specific bottleneck of running large models at acceptable latency. If correct, this positions XCENA between the GPU-first narrative and the inference-optimization layer above it.

What Might Happen Next Week

Two predictions, stated specifically enough that they can be checked.

Anthropic's Mythos model launches within two weeks and immediately resets coding benchmark expectations. Fortune's coverage of the Series H mentioned Mythos arriving "in coming weeks," alongside the Claude Opus 4.8 release. If Mythos ships in early June, it will be benchmarked directly against the tasks Cognition used to justify a $26B valuation. A base model capable enough to handle autonomous coding at high reliability narrows the product moat for any agent sitting on top of it. Watch SWE-bench results specifically for the signal.

The FTC's "merger in disguise" inquiry names a second deal before July. The Groq-Nvidia structure is the most visible example, but similar licensing-plus-executive-transfer arrangements appeared in at least two other AI transactions in the past 12 months. Groq's public $650M raise draws renewed attention to exactly the kind of company left behind after one of these deals. The agency does not move fast, but it does move publicly, and the next named inquiry is more likely to arrive before Q3 ends than after.