The Compute 100 — Thursday June 11, 2026

Anthropic's Fable 5, launched June 9, carries safety classifiers that permanently reroute biology, chemistry, cybersecurity, and AI hardware and kernel development queries to the less-capable Opus 4.8, while full Mythos 5 capability runs only through Project Glasswing — currently limited to US government cyberdefenders and critical infrastructure providers. On June 10, developer backlash arrived, see Fortune, while CEO Dario Amodei published a policy essay calling for government authority to block model deployments that fail independent safety testing — extending Fable 5's restriction logic into a proposed industry-wide regulatory mandate. The verticals first to face degraded capability are exactly those AI infrastructure builders run on: hardware development, model distillation, and security research; watch whether enterprise customers in these segments migrate to OpenAI or Google, and whether Amodei's proposals advance in Congress.

Amodei's June 10 essay "Policy on the AI Exponential" is worth the read to hear his perspectives on the regulatory rationale around the deployment of Mythos.

d-Matrix announced Corsair, its inference accelerator, has entered full production on TSMC N6 via Alchip Technologies, with volume shipments beginning to priority customers including Microsoft. The company claims Corsair running alongside Nvidia Blackwell delivers inference 10x faster, 3x cheaper, and 5x more energy efficiently than standalone GPU clusters. The latency requirements of tools like Claude Code continue to push inference beyond what GPU-only infrastructure handles efficiently. CEO Sid Sheth claims commitments from high-profile hyperscalers, neoclouds and frontier AI labs for the Corsair system.

WEKA and Oracle Cloud Infrastructure published benchmarks showing WEKA's NeuralMesh platform with Augmented Memory Grid delivers 10x token throughput, serves 10x more concurrent users, and produces 7x more tokens per GPU than DRAM-only configurations for long-context AI inference. The benchmark addresses a specific scaling constraint: as context windows lengthen toward 1M+ tokens, the storage and memory layer — not GPU compute — becomes the binding bottleneck for concurrent user capacity. The OCI deployment advances WEKA's spot in the inference infrastructure buying cycle.

Announced a $7 billion equity raise on June 9 after close — $1.25B common stock, $3.75B mandatory convertible preferred, $2B ATM program — to fund $39 billion in AI server orders from more than 20 customers. SMCI fell 28% on June 10 in a market already down nearly 2% (Nasdaq), with dilution risk driving the sell-off rather than any demand concern. The $39B backlog is real; the question is margins: server assembly is thin-margin, and mandatory convertible preferred creates ongoing dilution pressure unless the AI server cycle sustains at pace.

AI-related global debt issuance is on pace to top $570 billion in 2026, more than double last year, according to a Morgan Stanley forecast reported by Reuters on June 10. By May 31, hyperscalers, semiconductor companies, and data center developers had already raised $236 billion — four times the pace of the same period in 2025. For fixed-income investors, AI infrastructure is now the largest single allocation in investment-grade credit; the marginal actor setting cost of capital for the compute buildout is increasingly a bond portfolio manager, not a venture fund. That shift matters for how AI infrastructure capital is allocated and at what required return thresholds.

The UK government announced a £1.1 billion AI hardware strategy at London Tech Week: £750M for a national AI supercomputer targeting 2030 deployment, £150M for next-generation inference chips, £250M for specialized chips, £120M for an AI Hardware Innovation Programme supporting domestic chip design, and a new £150M venture fund backed by Playground Global and the British Business Bank. The plan explicitly targets British-designed chips for the national system. At current hyperscaler capex rates, £1.1B represents roughly one day of combined US hyperscaler spending — but as a chip sovereignty policy signal it follows the EU Chips Act and US CHIPS Act in treating domestic semiconductor capacity as a national infrastructure priority, not a commercial question.