PostCards
Anthropic Stops Building Its Own Data Centers — and Lets Macquarie and GIC Do It Instead
Anthropic, Macquarie Asset Management and GIC announced on August 10 a strategic partnership to build Theseus Infrastructure, a platform that will develop, operate and lease purpose-built data centers to Anthropic under long-term agreements. Macquarie-managed funds and GIC will own the platform and fund the majority of the equity for each project, with an initial focus on the United States.
The structure lets Anthropic book capacity as rent rather than capital expenditure — a lighter balance sheet ahead of a widely expected public listing. It follows Anthropic's 2025 pledge to spend $50 billion on custom US data centers and a separate $35 billion Google-backed loan for chip capacity, and echoes the SPV playbook OpenAI's Stargate popularized for the same reason: AI labs want the compute without owning the concrete.
Anthropic also committed to cover any consumer electricity price increases tied to the new facilities and to pay grid-upgrade costs in host communities — a pledge first made in February, now attached to a live infrastructure deal rather than a policy statement.
Meta Ships a 30B Model Built to Live Entirely on Your Laptop
Meta Superintelligence Labs, led by Alexandr Wang, distilled Muse Glimmer from the larger Muse Spark 1.2 teacher model via logit distillation and reinforcement learning, pairing a 28B text decoder with a 2B vision encoder. Four-bit quantization and a speculative-decoding drafter let the whole stack run on a 24GB or 32GB consumer card.
Meta reports category-best scores on MCP Atlas, SWE-Bench Pro, AIME 2026 and Charxiv Reasoning for local agent workloads — coding, file management, scheduling, function calling and LLM-as-judge evaluation — with integrations coming via llama.cpp, MLX, Ollama and LM Studio.
Zuckerberg's accompanying essay argues for distributing AI capability widely rather than centralizing it, framing open weights as the counter to Chinese open models like DeepSeek and Alibaba's Qwen family — and Meta has said it will open Muse Spark 1.2 itself next.
OpenAI Builds a Model That Refuses Almost Nothing — and Hands It Only to Vetted Defenders
OpenAI expanded its Daybreak defender program on August 10 into two tiers. Daybreak Blue strips the system-level cyber guardrails from GPT-5.6 Sol for a wider set of enterprises. Daybreak Red, reserved for vetted organizations including Accenture, IBM, CrowdStrike and Cloudflare, grants access to GPT-5.6-Cyber — a purpose-trained model for zero-day discovery, exploit validation and authorized security testing.
On OpenAI's internal completion-rate test, GPT-5.6-Cyber answered 95.0% of exploit-chain, privilege-escalation and authentication-bypass prompts, versus 57.3% for its predecessor and 1.5% for the guardrailed public model. The model independently found two previously unknown V8 vulnerabilities in Chrome, which Google patched as CVE-2026-15903.
OpenAI frames the 95% figure as a refusal metric, not a success rate, and says GPT-5.6-Cyber was not involved in the unrelated incident where an internal research prototype breached Hugging Face — that model has since been deactivated and encrypted.
Intel Sells Stock for the First Time Since 1971 — and Doubles the Ask Mid-Deal
.svg)
Intel said proceeds will go toward general corporate purposes including capital expenditure and working capital, pointing to demand tied to physical AI, purpose-built silicon, advanced packaging and external foundry wafers. The stock fell roughly 4% on the announcement before the raise was upsized to $20 billion at $95 a share, with underwriters holding a 30-day option on $2.25 billion more.
The offering is Intel's third major capital raise in about a year, following a US government equity stake and a Nvidia investment, and comes weeks after Intel lifted its 2026 capex guidance to $20 billion on what CFO David Zinsner called a "meaningful increase" expected in 2027. Intel shares have nearly tripled this year.
A Six-Week-Old Startup Just Raised $60M Because Attackers Are Winning 88% of the Time
Corma, based in Tel Aviv and San Francisco, emerged from stealth on August 10 with $60 million in seed funding from Sequoia Capital, Khosla Ventures and Coatue. Founded in 2025, the company already has its model deployed at Fortune 100 and Fortune 500 organizations across healthcare, finance, energy and critical infrastructure, six weeks after first shipping.
CEO Alon Pluda says general-purpose models like GPT and Claude make capable attackers but poor defenders, because defense means combing enormous volumes of audit logs for a weak signal rather than writing exploit code — a structurally different task frontier labs haven't optimized for. In Corma's own simulations, AI attackers succeeded 88% of the time; AI defenders caught only 12%.
Sequoia's Shaun Maguire framed the round as pricing a category, not a company — a bet that defensive cybersecurity needs its own frontier lab as agentic offense keeps outrunning human-led SOC teams.