THE SIGNAL
05 — 11 MAY 2026
FRONTIER TOWER04

On May 6th, Anthropic announced it had secured all of the compute capacity at SpaceX's Colossus 1 data center — over 300 megawatts and more than 220,000 NVIDIA GPUs deployable inside a month. The deal also names "multiple gigawatts of orbital AI compute capacity" as the next phase, alongside existing partnerships with Amazon (5 GW), Google (5 GW), Microsoft and NVIDIA ($30 billion on Azure), and Fluidstack ($50 billion). Rate limits on Claude Code doubled the same day across Pro, Max, Team, and Enterprise plans; peak-hour throttling was removed from Pro and Max.

The mechanism is the calendar. On May 1st the Pentagon excluded Anthropic from its classified-network AI contracts, having designated the lab a supply-chain risk after the company refused to drop usage-policy restrictions on military targeting. On May 4th Anthropic shipped a $1.5-billion joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs. On May 6th, SpaceX. Three days. One classified procurement lost, one finance channel built, one piece of compute infrastructure the size of the largest hyperscale build in industry history acquired. The pace is the signal — the company is not negotiating, it is buying.

What lands in the room with 300 megawatts is the trade. The frontier through 2024 and 2025 was a model trade — whose Opus, whose GPT, whose Gemini. The frontier in May 2026 is a compute trade — whose 300 megawatts, whose 220,000 GPUs, whose orbital ambition. The Pentagon exclusion looked like a setback at the start of the week. By Wednesday it read as the moment Anthropic stopped renting the frontier and started owning the substrate underneath it. The classified contracts will be back at the table when the only company that can serve them at scale is the one that bought the Colossus.

On May 6th, OpenAI published MRC — Multipath Reliable Connection — a new RDMA transport protocol that distributes a single connection's traffic across multiple network paths in parallel. NVIDIA's Spectrum-X is the proving hardware. The protocol features microsecond-level failure detection and rerouting in silicon, eliminating the GPU-idle window that conventional RDMA single-path transports incur during a fabric fault. The spec was released as an Open Compute Project submission and developed in concert with AMD, Broadcom, Intel, and the major cloud providers. MRC is already in production at Oracle and Microsoft.

The structural detail is what the protocol replaces. Conventional InfiniBand and Ethernet RDMA bind one connection to one path. A single congested link or a single failed switch port stalls the entire frontier training run. MRC unbinds the connection from the path — traffic load-balances across every available route, hardware-accelerated retransmission recovers from packet loss, and a path failure triggers a sub-millisecond rebound rather than a model checkpoint reset. OpenAI's Sachin Katti — *"MRC's end-to-end approach enabled us to avoid much of the typical network-related slowdowns and interruptions and maintain the efficiency of frontier training runs at scale."* The unspoken half of the quote is that frontier-run interruptions used to cost six and seven figures per incident.

The implication is that the network is now a model. The protocol authors — OpenAI, NVIDIA, AMD, Broadcom, Intel — just shipped what amounts to a new IETF stack purpose-built for hundred-thousand-GPU runs. The OCP submission makes it open to anyone who can run the fabric. The cloud providers (Oracle, Microsoft) are already paying customers. For builders, MRC is not optional reading; the next eighteen months of frontier training will be conducted on it, the next eighteen months of inference fabric will be deployed against it, and the gap between a lab running on legacy RDMA and one running on MRC is the gap between a six-figure-per-hour idle GPU bill and zero.

On May 5th, Cerebras Systems priced its U.S. IPO range at $115 to $125 per share, targeting a fully diluted valuation around $26.6 billion. The company manufactures the wafer-scale WSE-3 chip — a single piece of silicon the size of a dinner plate that contains the entire training fabric of a conventional GPU cluster on one die. The pricing arrived the same week Anthropic took 300 megawatts at Colossus 1 and OpenAI and NVIDIA shipped MRC — three compute-side events in the same calendar week.

The structural fact is that Cerebras goes public into the largest compute demand market in the history of computing. NVIDIA's data-center revenue line was $130 billion in fiscal 2025. Google Cloud's AI backlog nearly doubled to $460 billion in a single quarter, per Francis deSouza at Milken. The bottleneck is not a software bottleneck. It is a wafer bottleneck, a fab bottleneck, an ASML EUV-scanner bottleneck. Cerebras's pitch — skip the cluster, ship the wafer — is the bet that single-die training fabric beats networked GPU fabric for a defensible slice of the frontier-training workload.

The implication for builders is asymmetric. NVIDIA's H100, H200, and Blackwell SKUs are sold out into 2027. Cerebras is the only public-market entry into the wafer-scale alternative. The IPO price prices the alternative at roughly one-tenth of NVIDIA's data-center pure-play valuation — a discount that reflects the smaller installed base and the longer software-stack lead time, not the underlying technology bet. Whether the alternative is mispriced is the next two quarters' question. The chip economy has its second listed AI-native pure-play; the first one has trillion-dollar market-cap inertia and a public-market exit route that other vendors do not.

On May 7th, OpenAI published the operational rollout of Trusted Access for Cyber — the partner program announced at the end of last week's "Cybersecurity in the Intelligence Age" post. Both GPT-5.5 and GPT-5.5-Cyber are now available to verified defenders, alongside accelerated vulnerability research and protection workflows. The framing — the same model the UK AI Security Institute scored at 71.4% on advanced expert cyber tasks is now distributed under terms that look structurally identical to Anthropic's Project Glasswing.

The mechanism is verification at the customer layer. Public-tier ChatGPT users get the standard GPT-5.5 capabilities. Verified defenders — meaning enterprise customers with attested cybersecurity functions — get access to the cyber-tuned variant plus tooling, plus the implicit understanding that capability deltas the public never sees are routed through this tier. The contrast with last quarter's posture is the news. In Q1, OpenAI executives publicly criticized Anthropic for gating Mythos. In Q2, OpenAI shipped the same gate.

The watchlist is the next-twelve-months question. Who else gets a tier? Which capabilities land where? The next vector to watch is the defense and intelligence channel for whom Trusted Access is the rehearsal — the labs that gated Cyber will gate the next capability, and the gate will be the procurement contract.

On May 5th, Anthropic published ten ready-to-run agent templates aimed at the financial-services workflow — pitch builder, meeting preparer, earnings reviewer, model builder, market researcher, valuation reviewer, general-ledger reconciler, month-end closer, statement auditor, KYC screener. All ten run on Claude Opus 4.7, which scores 64.37% on [Vals AI's Finance Agent benchmark](https://vals.ai/). Customers in the launch card — Citadel, FIS, BNY, Carlyle, Mizuho, Travelers, Walleye Capital, Hg. Quote of the launch is from Mizuho's Patrick Suehnholz — *"Claude compresses and enhances the work before the meeting."*

The mechanism is the connector inventory. Eight new data integrations land with the agents — Dun & Bradstreet, Fiscal AI, Financial Modeling Prep, Guidepoint, IBISWorld, SS&C Intralinks, Third Bridge, Verisk. On top of that, Moody's shipped an MCP app the same day exposing credit ratings on more than 600 million companies. The agents deploy three ways — plugins inside Claude Cowork or Claude Code; Claude Managed Agents cookbooks for autonomous runs on the Claude Platform; or as add-ins for Microsoft 365 Excel, PowerPoint, and Word.

The implication is that the Blackstone JV was the channel, and this is the inventory. Last week's beat was the announcement of an enterprise services company with three Wall Street partners. This week's beat is the first ten products that company will resell. The procurement officer at a $20-billion private-equity firm now does not buy "a Claude license." They buy "the KYC screener" or "the earnings reviewer" — finished workflows with a benchmark score, a customer card, and a data-connector stack that ships with the agent.

On May 7th, Google DeepMind published the eight-month update on AlphaEvolve — its Gemini-powered coding agent. The receipts are the story. Klarna — transformer training throughput doubled. FM Logistic — 10.4% routing-efficiency gain, 15,000 kilometres of truck driving saved annually. WPP — 10% accuracy gain in campaign optimization. Schrödinger — 4x speedup in molecular force-field training. Google's own Spanner — 20% reduction in write amplification. Compiler optimization — 9% storage-footprint reduction. Cache replacement — months of optimization done in two days.

The mechanism is evolutionary search guided by Gemini. AlphaEvolve does not generate code top-down from a spec the way Claude Code or Codex does. It mutates a candidate solution, runs the candidate against an evaluation function, and iterates. The category targets — quantum-circuit error rates 10x lower than baseline; new bounds on the Traveling Salesman Problem; record-breaking solutions to Ramsey Numbers; the DeepConsensus DNA-variant detection that cut variant-call errors by 30%. The eight-month track record is the test the recursive-self-improvement thesis is going to be evaluated against from here.

The implication is that the agent infrastructure story that filled last issue's beat IV just got its first set of industrial-deployment numbers. The line that survives the dispatch — when an AI coding agent ships into Spanner, the database that holds Google Search's index, the deployment is past the demo phase. AlphaEvolve is a category — *evolutionary coding agent* — not a brand. The next eighteen months will be the period in which every cloud vendor ships their version of it, and the procurement question will be which one's evaluation function is best instrumented.

On May 5th, Anthropic co-founder Jack Clark published Import AI 455 — "AI systems are about to start building themselves." Headline claim — Clark assigns 60%+ probability to no-human-in-the-loop AI R&D occurring by end of 2028. The mechanism — proof-of-concept in 2026 at non-frontier scale; 30% probability by end of 2027 (limited by creativity); 60%+ by end of 2028 (a *Rubicon into a nearly-impossible-to-forecast future*). On May 7th, Matthew Hutson at IEEE Spectrum filed the survey piece — "AI Is Starting to Build Better AI" — with named systems and 25 expert interviews.

The mechanism is the existing track record. Clark cites the SWE-Bench progression — 2% in 2023, 93.9% on Mythos Preview in 2026 — and the time-horizon expansion from 30-second tasks in 2022 to 12-hour reliable task completion in 2026. Hutson's piece names the systems already in the loop — OpenAI's GPT-5.3-Codex helping with its own training; AlphaEvolve optimizing chip design at Google; Darwin Gödel Machines from UBC and Sakana AI; the AI Scientist paper in Nature in March. Of 25 AI experts surveyed by Hutson, most thought labs would keep recursive-improvement models internal.

The implication for the dispatch is the wager. Three issues ago the question was whether the frontier doubles each quarter. Today the question is whether the frontier folds in on itself — whether the model that ships in Q3 2027 was authored by the model that shipped in Q1 2027. Clark's 60% number is one lab's prior; the IEEE Spectrum survey is the institutional check. The right posture for a builder is neither *it's already happening* nor *it's hype.* The right posture is to track the milestones, stake the dates, and re-mark when each one passes — because the moment the loop closes, the dispatch the week before will be the last one written entirely by humans about humans.

On May 6th, five architects of the AI economy sat on a Milken Institute Global Conference panel in Beverly Hills — Christophe Fouquet (ASML CEO), Francis deSouza (Google Cloud COO), Qasar Younis (Applied Intuition co-founder), Dmitry Shevelenko (Perplexity chief business officer), and Eve Bodnia (Logical Intelligence founder). TechCrunch filed the recap. The headline number was deSouza's — *Google Cloud's AI backlog nearly doubled to $460 billion in a single quarter.* Fouquet's posture — the chip supply will remain *"supply limited"* for the next two, three, maybe five years.

The mechanism is the constraint stack. Fouquet from ASML — EUV scanners are the bottleneck and the cycle to install one is multi-year. deSouza from Google Cloud — the demand backlog grew faster than the supply could fill, and orbital data centres are now under exploration because the terrestrial-power ceiling and the local-grid permitting cycle are both binding constraints. Bodnia from Logical Intelligence — the energy-based models thesis directly challenges whether LLM-shaped architectures are the optimal path forward at all, raising the question of whether the $460-billion backlog is a backlog for the *right* product. Younis from Applied Intuition — physical AI raises sovereignty questions (national procurement of foreign-built autonomous systems) that the cloud-AI procurement template does not.

The implication for the dispatch is that the bull case and the wheels-coming-off case are the same case. The reason the backlog doubled is the reason ASML cannot supply the scanners; the reason orbital data centres are in slides is the reason Anthropic just bought 300 megawatts at Colossus; the reason Bodnia is asking the architecture question is the reason a Cerebras IPO lands the same week. The AI economy is operating at the limit of its physical inputs. Whether the inputs come from a fab in Phoenix, a megawatt in Colossus 1, or a satellite tier nobody has built yet — that question is the entire next eighteen months.

On May 8th, Nextgov reported a senior Department of Defense official stating the department will *"never again"* rely on a single AI provider. The statement lands one week after the May 1 classified-network contracts went to OpenAI, Google, NVIDIA, Microsoft, AWS, SpaceX, and Reflection AI — with Anthropic conspicuously excluded. The same week Anthropic took 300 megawatts at Colossus 1.

The mechanism is the policy framing. The May 1 contracts read as the department picking five-plus-two labs and excluding the safety-frame holdout. The May 8 walk-back reframes the procurement — multi-vendor is now doctrine, not preference, and the exclusion of any single lab is no longer permissible as a structural posture. The unstated half is the Anthropic JV announcement on May 4 (Wall Street as alternative channel) and the SpaceX deal on May 6 (substrate ownership). The Pentagon is signalling that the door is open while the procurement team rewrites the policy that closed it.

The reckoning is the leverage. The Pentagon walked the single-provider posture back not because the safety-frame argument got more persuasive, but because the $900-billion company the department excluded just acquired the compute substrate the department's classified workloads will eventually need. The May 1 exclusion looked decisive at filing. By May 8 it reads as a policy moment that will be undone the moment a classified workload requires a model that runs on Claude. The lab that was supposed to be the cautionary tale is the lab whose absence rewrote the procurement doctrine in seven days.

On May 6th, Apple agreed to pay $250 million to settle a class-action lawsuit over Apple Intelligence's Siri features — the upgrade announced at WWDC June 2024 and never shipped. Eligible class — U.S. customers who purchased an iPhone 15 or iPhone 16 between June 10, 2024 and March 29, 2025. Individual payouts run up to $95 per device. Apple did not admit wrongdoing. The plaintiffs argued Apple's marketing created the impression advanced AI capabilities would arrive sooner than they actually did.

The mechanism is the lag. Apple shipped Apple Intelligence at the rumored functionality in the iPhone 15 Pro launch cycle and the iPhone 16 base; the enhanced Siri that would carry contextual conversation, on-device agentic work, and a ChatGPT-or-Claude-class voice assistant was the marketing pillar that did not materialize on the original cadence. Apple has reportedly targeted WWDC 2026 on June 8 for the enhanced Siri reveal. The settlement caps the cost of the missed cycle; the WWDC reset is now legally enforced.

The implication is the trust ratchet. Twenty-five years of Apple product marketing carried a presumption that announced features ship on schedule. Apple Intelligence broke that presumption in the most visible product the company sells, and the courts just put a number on the cost. The number per device — $95 — is small. The structural lesson — that *announcing* a model feature is no longer free, and that the Cupertino AI program is operating under an enforcement posture the rest of the industry has not yet faced — is the receipt. June 8 is the next data point. Anyone whose dashboard tracks AI competitive position should bookmark WWDC.