#01 - OpenAI Builds Its Own Chip, and Names It Jalapeño
On June 24 OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom inference chip, built from scratch in nine months with Celestica. It is purpose-built for LLM inference, not training, and early tests show substantially better performance per watt than current hardware. First deployment lands end of 2026 at gigawatt scale with Microsoft.
Why it matters: Custom silicon for inference signals where the economics are heading. Expect inference costs to fall as this scales, but also expect OpenAI's vertical integration to squeeze third-party infra providers competing on price. If you run at scale, the cost curve is about to bend.
#02 - Anthropic Accuses Alibaba of Industrial-Scale Claude Extraction
In a June 10 letter to the US Senate, Anthropic accused Alibaba's Qwen lab of the largest known distillation attack against Claude to date: roughly 25,000 fraudulent accounts generating 28.8 million interactions between April and June, targeting coding and agentic reasoning capabilities. It is a lobbying move toward tighter rules, not a lawsuit. Alibaba has not publicly responded.
Why it matters: API access is becoming legally and politically contested territory. If you resell or integrate frontier models, expect tighter access controls and account verification coming down the pipe. The era of frictionless API access is closing.
#03 - OpenAI May Push Its IPO to 2027 to Chase a Trillion
Weeks after filing confidentially, OpenAI is reportedly weighing a delay of its IPO to 2027, per the New York Times and Bloomberg. Sam Altman is said to prefer holding out for a valuation near 1 trillion dollars rather than listing into 2026's shaky market. SpaceX's post-IPO slide reportedly made advisors cautious. Nothing is confirmed.
Why it matters: For founders benchmarking against OpenAI's valuation, a delay extends the fog of private-market pricing. It also hints that even OpenAI's leadership is not yet confident in the public-market story, which is a signal worth reading.
[A MESSAGE FROM CONCEPT OF PROOF]
We build working AI tools for real business problems. No decks, no advice, just a running system in weeks. See what we have shipped at conceptofproof.nl
#04 - Chinese Models Are Closing the Gap, Fast
Multiple reports indicate Chinese labs, notably Zhipu AI, have reached parity with Anthropic and OpenAI on key benchmarks including cybersecurity, even as US export controls constrain how freely American labs can ship. The assumption that frontier capability is a US-only advantage is eroding.
Why it matters: If you choose a model provider for a long-lived product, factor in that Chinese alternatives are becoming viable and increasingly accessible outside the US. Provider risk now includes geopolitics, not just uptime.
#05 - Anthropic's Economic Index Maps Where AI Actually Sticks
Anthropic published its Economic Index, primary research measuring how Claude is actually used at work and where it substitutes for versus augments human labor. It identifies recurring task structures called Cadences that reveal where AI adoption is taking hold in professional settings.
Why it matters: If you design AI-powered workflows, this is empirical grounding on where adoption actually sticks. Worth reading before you assume which use cases drive retention rather than churn.
#06 - The Fable Ban Drags On, and Open Weights Win
Three weeks after the US ordered Anthropic to disable Claude Fable 5 and Mythos 5, the saga keeps shifting: parts of Mythos reportedly cleared for limited release, while the models stayed dark for most users. In the gap, Chinese open-weight models pitched themselves hard on one line: weights you own cannot be recalled by a government.
Why it matters: This is the throughline of the past month. If your product rides a single frontier API, a directive can vanish your dependency overnight. Self-hosting is no longer ideological, it is a continuity plan.