Morning newsletter digest

Axios Pro Rata: “OpenAdvisory”

• OpenAI launched The OpenAI Deployment Co., or DeployCo, a consulting and services arm backed by $4B at a $10B pre-money valuation, with OpenAI retaining majority control.
• Bain, Capgemini, and McKinsey are investors, which is either smart proximity to the frontier model roadmap or a tidy way to help fund their own disintermediation.
• Architect Capital is investing $535M into OnlyFans for about 16%, implying a $3.15B valuation despite $684M in 2024 pre-tax profit.
• Why it matters: AI deployment is becoming a PE-backed services race, not just a model race. Watch whether KKR, Carlyle, and EQT join the lab-affiliated services pile-on.

Stratechery: “The Inference Shift”

• Ben Thompson argues the market is splitting into three compute categories: training, answer inference, and agentic inference. Nvidia remains structurally advantaged in training, while fast answer inference creates room for systems like Cerebras and Groq.
• The real shift is agentic inference. If agents are doing work without a human waiting on every token, latency matters less and memory, state, history, KV cache, databases, logs, embeddings, and object stores matter more.
• That implies a different architecture: less premium for bleeding-edge GPU speed and HBM everywhere, more value in memory hierarchies with cheaper, higher-capacity memory and “good enough” compute.
• Why it matters: This is the day’s best strategic read. The bullish AI compute story may broaden from “buy the fastest Nvidia stack” to “who can orchestrate cheap, reliable, memory-rich infrastructure for autonomous work.” Very All-In coded, in the best way: market structure before ticker worship.

The Rundown AI: “Google DeepMind’s powerful AI co-mathematician”

• DeepMind published an AI co-mathematician system built around Gemini 3.1, modeled more like a coding-agent environment than a single-answer chatbot.
• It uses a coordinator plus sub-agents for parallel proof attempts, literature search, code, and review.
• It reportedly hit 48% on Epoch AI’s FrontierMath Tier 4, more than doubling Gemini 3.1 Pro’s 19% raw score, and helped Oxford’s Marc Lackenby spot a useful proof strategy inside a rejected output.
• Why it matters: The useful framing is not “AI replaces mathematicians.” It is “agentic workflows turn frontier models into research accelerants.”

Fortune Term Sheet: “The bad phishing email is dead”

• Frame Security launched publicly with $50M from Index Ventures, Team8, Picture Capital, plus participation from Wiz CEO Assaf Rappaport and Elad Gil.
• The thesis: AI has made phishing cheaper, more personalized, and more convincing, so employee-focused security is becoming a real category again.
• CEO Tal Shlomo says the world of badly written phishing emails is basically gone.
• Why it matters: Human-risk security is getting venture validation because the attack surface is no longer just inbox spam. It is cloned voices, believable texts, and context-rich impersonation.

Short Squeez: “Latest 6,000% Meme Stock”

• AST SpaceMobile has become a meme-stock phenomenon, up nearly 6,000% over 22 months at its peak and reaching a $25B valuation on just $71M in annual revenue.
• The bull case is real enough: space-based cell coverage, deals with AT&T, Verizon, Vodafone, and Alphabet as a major shareholder.
• The bear case is also wearing a neon sign: SpaceX, Amazon, massive satellite deployment needs, and a valuation already pricing in a very big win.
• Why it matters: Retail speculation is alive and well. The market can still turn a plausible infrastructure story into a rocket ship with a cult following strapped to the side.

Worth opening: Stratechery’s “The Inference Shift” first, then Axios Pro Rata’s OpenAI DeployCo piece.