THE SIGNAL
31 MAY — 6 JUN 2026
FRONTIER TOWER08

On June 1st, Anthropic confidentially submitted a draft registration statement on Form S-1 to the SEC — six days after closing the $65 billion Series H that pushed its post-money valuation to $965 billion. The funding arc became an IPO track in under a week, and in doing so the lab vaulted ahead of OpenAI, which has not yet reported filing any paperwork of its own. The new fact is not the money — that was last week — it is the filing, and the order it puts the race in.

A confidential draft S-1 is a registration statement, not an offering. By Anthropic's own words, *"the number of shares to be offered and the price have not yet been set"* — no share count, no per-share range, no exchange, and a go-public option that only opens after the SEC completes its review. What the filing does ratify is the revenue underneath it: Anthropic now reports a $47 billion annualized run-rate, up from roughly $9 billion at the end of 2025 — a five-fold climb in two quarters, fueled by Claude Code and the enterprise demand that carried it past OpenAI on business adoption in April.

What goes to the SEC is a number with a witness. The $965 billion was underwritten by seven private funds in a single round; the S-1 invites the public markets to underwrite it instead, on audited financials and a price the company has pointedly declined to name. The filing converts a valuation that asked to be trusted into one that asks to be cleared — and it does so the same week DeepSeek opened its first-ever outside round, the capital order of the frontier reshuffling on both sides of the Pacific at once. The lab that priced itself for public markets has set everything except the price.

Three days after confidentially filing a draft S-1 with the SEC — the paperwork that puts a $965 billion lab on a public-market track — Anthropic published **When AI builds itself** and told the world to consider stopping. The June 4th report, co-authored by Marina Favaro and Jack Clark, argues "it would be good for the world to have the *option* to slow or temporarily pause frontier AI development." The same week it asked the market to underwrite the race, it asked the field to brake.

The mechanism Anthropic is afraid of is **recursive self-improvement** — an AI system "capable of fully autonomously designing and developing its own successor," the point where the loop closes and humans stop driving each step. The report's own ledger says the loop is already partway internal: as of May 2026, more than 80% of the code Anthropic merges into its codebase was written by Claude, up from low single digits before Claude Code shipped in February 2025. The proof of the danger is the company's own commit history.

The ask is conditional, and the condition is the whole problem. A unilateral pause, Anthropic concedes, only works if everyone pauses — "if a slowdown simply lets the least cautious actors catch up technologically, it could leave everyone less safe." So the most valuable lab on Earth has filed to sell shares in a capability it argues is too dangerous to keep racing on, and proposed a brake that nobody can pull alone. The S-1 asks the market to price the upside; the report asks the market to fund the thing it just warned about.

On June 3rd, Reuters reported, via the South China Morning Post — citing people who declined to be named — that DeepSeek is raising roughly $7.4 billion (about 50 billion yuan) in its first-ever round of outside money, at a post-money valuation of $52 to $59 billion. Tencent is weighing about 10 billion yuan and battery maker CATL about 5 billion yuan, with fewer than ten strategic investors in the deal. Founder Liang Wenfeng is putting in roughly 20 billion yuan of his own capital — about 40 percent of the round, the dominant individual stake. The terms are not closed; the round is expected to finalize within weeks, and the numbers could move.

The reversal is the mechanism. DeepSeek built its reputation refusing this exact transaction — Liang self-funded the lab out of his quant-fund fortune precisely so it could chase artificial general intelligence without a board demanding returns, and shipped V3 and R1 as open-weight models that undercut the closed frontier on cost. That posture is now meeting its compute bill. The reported use of proceeds is not a pivot to product: it is larger GPU clusters, more semiconductors, and richer pay to hold talent. The valuation jumped roughly six-fold from the $10 billion mark in April — the market repricing the lab the moment it agreed to be priced at all.

The implication is that cheap has a capital floor. DeepSeek's whole thesis was that you could reach the frontier without the hyperscaler war chest — and in the same week Anthropic filed to go public at a near-trillion-dollar valuation, China's champion conceded that staying cheap at the frontier still costs billions. The open-weights, low-margin strategy does not escape the compute arms race; it just delays the invoice. When the lab that proved frontier AI could be self-funded takes Tencent's and CATL's money to keep going, the signal is not that DeepSeek got greedy. It is that the floor under the whole race rose, and even the cheapest player on the board can no longer stand on it alone.

On June 3rd, Berkeley's RDI lab released Agents' Last Exam, a benchmark built with 250+ industry experts across 13 industry clusters and 55 subfields — real, economically-valuable work in After Effects, Siemens NX, Unreal Engine, and Rhino 3D, anchored to the federal O*NET / SOC 2018 occupational taxonomy. On the hardest tier, the frontier's average full pass rate is **2.6%**. Not a leaked eval, not a saturated one — a fresh measurement of whether agents can do the jobs they are sold as replacing.

The cruelty is in the spread. The strongest configuration, **Codex** on **GPT-5.5** — which posts 82% on Terminal-Bench — scores below 50% on the *easiest* tier and 8.6% on the hardest. Claude Code lands at 0.0% on that hardest tier. These are pass/fail on full task workflows, not partial-credit token overlap — an **unsaturated benchmark**) that, unlike the SWE-bench numbers in every launch deck, nobody has had time to overfit, contaminate, or game. The corpus is 1,500+ tasks toward a 5,000 target; the paper's frozen subset reads 1,490, and the live site is newer.

This is Beat II's ledger entry against Beat I's $965B. The same week the frontier filed to go public on the premise that the work is nearly done, a clean instrument said the work is barely started — 90%-on-SWE-bench does not survive contact with After Effects. The contamination arc that ran through week-07 closes here: the question is no longer whether the benchmarks are dirty. It is whether the frontier can do real economic work at all, measured clean. The answer, for now, is 2.6%.

On June 1st, MiniMax shipped M3, an open-weight coding model with a 1-million-token context window, and the pitch was a straight shot at the frontier: **59.0%** on SWE-Bench Pro, surpassing GPT-5.5 and Gemini 3.1 Pro and approaching Opus 4.7, plus 66.0% on Terminal-Bench 2.1 — at a reported five to ten percent of the cost of the closed labs it names. The API went live the same day. The model that earned the number did not.

Every one of those figures is company-reported. MiniMax states the SWE-Bench Pro run was "tested on internal infrastructure using Claude Code as the scaffolding" — its own machines, its own harness, no independent rerun — and the-decoder notes the results carry no outside verification. The cost edge is real engineering: **MSA**, MiniMax Sparse Attention, drops per-token compute to one-twentieth of the prior generation at a million tokens. But "open-weight" was the part that made the claim auditable, and at launch the weights were not released — promised on Hugging Face and GitHub "within the next ten days." A week later they are still not out: the GitHub repository's own README reads, "The model is not yet released," and points users back to M2.7.

So the receipt is a number nobody outside the building can reproduce, attached to a model nobody outside the building can run. That is not fraud; it is the new default. In a season when vendor-graded coding scores have quietly stopped meaning much — see this issue's Agents' Last Exam, the clean instrument that put the frontier at 2.6% — MiniMax launched on a self-run benchmark and led with it. For a builder, the lesson is the asterisk, not the leaderboard: a frontier-parity claim with no chain of custody is a marketing artifact until the weights drop and someone else runs the eval. Ship the checkpoint, then we will talk about 59.

At Build 2026 on June 3rd, Microsoft took the Agent Framework harness to 1.0 GA — the layer everyone rebuilds by hand. Shell access, filesystem access, human-in-the-loop approval, and context management across long-running sessions, all shipped as supported primitives. Automatic context compaction, a `FileMemoryProvider` for session memory, and a `TodoProvider` for task tracking now come in the box. The same day, GitHub Copilot SDK integration reached 1.0 alongside it.

The sharper move ships in the alpha `agent-framework-hyperlight` package: CodeAct. Instead of emitting one tool call per turn, the model writes a single short Python program that invokes your tools through `call_tool(…)`, and that program runs in a fresh, locally isolated Hyperlight micro-VM per call. On a representative multi-step workload — order totals across many users, dozens of tool calls — the traditional loop took 27.81 seconds and 6,890 tokens; CodeAct took 13.23 seconds and 2,489. That is a 52.4% cut in latency and a 63.9% cut in tokens, the model writing code to orchestrate the work it used to narrate one call at a time.

The hand-rolled agent loop was a moat made of glue. Microsoft just commoditized the glue, and then argued the loop itself is the wrong abstraction — that for tool-heavy work the model should write the program, not play the dispatcher. Run `create_harness_agent` and you inherit the compaction, the memory, and the todo tracking you were about to write yourself. The question stops being whether you can build the runtime, and becomes whether the runtime you built is faster than the one that now ships with a version number.

On June 2nd, Cognition retired the Windsurf brand and shipped what replaces it: Devin Desktop, pushed to existing users as a standard over-the-air update. The editor is no longer the center of the IDE. The new default surface is the Agent Command Center — a Kanban board where, in Cognition's words, "you manage every local and cloud agent from a single Kanban view." Days after the company raised $1 billion at a $26 billion valuation, it spent the product launch arguing the editor is the wrong abstraction.

Two pieces make the shift load-bearing. The local agent was rewritten from scratch in Rust as Devin Local, which Cognition claims is up to 30% more token efficient than the Cascade engine it retires; legacy Cascade runs only through July 1st. The interop layer is the Agent Client Protocol, an Apache 2.0 standard for editor-to-agent communication. Devin Desktop speaks it, so Codex, Claude Agent, OpenCode, and any other ACP-compatible agent run as first-class citizens beside Devin, with a primitive called Spaces letting related agents share context.

The agent loop most builders hand-roll — spawn three vendors' CLIs in three terminals, reconcile their diffs by hand — shipped this week as the default chrome of a shipping IDE. ACP is the part that matters: an open protocol means the host stops being a vendor lock. Adopt it and your in-house agent plugs into any compliant editor; refuse to pick one vendor's agent and run them side by side instead. The dev tool's unit of work just moved from the file you edit to the fleet you dispatch.

At Build on June 2nd, Microsoft launched seven in-house MAI models — MAI-Thinking-1, MAI-Code-1-Flash, MAI-Image-2.5 and its Flash variant, MAI-Transcribe-1.5, and MAI-Voice-2 and its Flash variant — and CNBC read the move plainly: the lineup exists to lessen reliance on OpenAI and lower costs for developers. After investing $13 billion in the partner whose models it has resold since 2023, Microsoft AI chief Mustafa Suleyman said the in-house stack, tuned on consulting-firm workloads, outperformed GPT-5.5 at ten times better cost efficiency. MAI-Code-1-Flash is already live in GitHub Copilot and VS Code, in the model picker and under the default auto router.

The receipts are company-reported, and worth reading as such. Microsoft says MAI-Code-1-Flash — a 5-billion-active-parameter agentic coder — beats Claude Haiku 4.5 on SWE-Bench Pro, 51.2% to 35.2%, a 16-point lead, while solving harder problems with up to 60% fewer tokens on SWE-Bench Verified. The reasoning model, MAI-Thinking-1, is a 35-billion-active, roughly one-trillion-total mixture-of-experts trained from scratch on licensed data with no distillation from any third-party model; Microsoft reports it matches Opus 4.6 on SWE-Bench Pro and posts 97.0% on AIME-2025. The pitch is not that these win the absolute frontier — it is that Microsoft owns the weights, runs them on Azure, pays no partner royalty, and hands the saving down the stack.

The geometry is the story. In the same news cycle, OpenAI's GPT-5.5, GPT-5.4, and Codex went generally available on Amazon Bedrock on June 1st — first-party pricing, usage counting toward existing AWS commitments, the first time those models shipped on a non-Microsoft cloud. So OpenAI now sells through Microsoft's largest rival while Microsoft ships models built to need OpenAI less: the most consequential partnership in the industry quietly un-bundled from both ends at once. For a builder, the abstraction that loosened this week is the assumption that the cloud and the model travel together. The frontier model is becoming a commodity you route to on cost, and the lab and the cloud you buy it from no longer have to be the same company.

On June 5th, a SpaceX free-writing prospectus disclosed that Google agreed to pay SpaceX $920 million a month for compute capacity, running October 2026 through June 2029. The package is approximately 110,000 NVIDIA GPUs, CPUs, memory, and related silicon, sited at the same xAI data centers SpaceX already leases to Anthropic. Filed Pursuant to Rule 433 against an active S-1, the line landed in the same week SpaceX itself filed to go public — the landlord priced the floor on the way to the auction.

The mechanism is that SpaceX owns neither the chips nor the breakthrough — it owns the power and the buildings around Colossus near Memphis, and it rents that substrate to whoever is short. Google called it "bridge capacity to meet surging customer demand" for Gemini Enterprise; the agreement ramps through September at a reduced fee, with a hard GPU-delivery deadline of September 30th before Google can walk. At $11 billion a year, Tom's Hardware notes SpaceX's projected data-center revenue would top its Starlink, launch, and AI proceeds combined.

This is the second SpaceX landlord deal in a single quarter, following the $1.25-billion-a-month Anthropic lease that week-06 read as the moment the limit became electrons, not chips. The renter is no longer paying for a model; it is paying for a substation, a slab, and a delivery date. Owning electrons-plus-silicon is the franchise now — and the quietest detail is that Musk's xAI is Google's compute landlord while Musk litigates OpenAI.

On June 1st, on the GTC Taipei stage at Computex, Jensen Huang announced the **Vera Rubin** platform "now in production" — Nvidia's next rack-scale generation, purpose-built for what he called agentic AI factories. The keynote ran a single thesis on a loop: agentic AI has arrived, it works, and it pays. Underneath the slideware sat a number with no marketing in it. Five years ago Nvidia spent between $10 billion and $15 billion a year in Taiwan; that figure is now roughly **$150 billion**, with AMD pledging over $10 billion into the same supply base over a comparable window.

The mechanism is the Vera Rubin NVL72, and the load-bearing detail is manufacturability, not FLOPS. Its cable-free, hose-free, fanless modular tray cuts assembly "from two hours to five minutes per compute tray" — Nvidia is optimizing the build line the way it once optimized the die, because the binding constraint on an AI factory is no longer the chip but how fast you can rack, power, and cool it. Huang made the accounting explicit — compute is revenue, every token produced profitable — which reframes performance-per-watt, reliability, and system lifetime as the financial levers, not the spec sheet.

That is the supply side of the franchise THE SIGNAL has tracked all quarter, and it closes a loop. The demand side rents — Google is paying SpaceX about $920 million a month for GPU capacity — while the supply side builds the silicon both of them rent. For a builder, the repricing is the signal: when the company that makes the picks declares every token a billable unit and spends $150 billion a year to mint more of them, inference stops being a cost line and becomes the meter. The frontier is no longer selling computers. It is selling the throughput, and it has started counting it by the token.