The AI story split in two this week. On the public markets, the trade that has powered equities all year finally wobbled — a broad sell-off in AI and chip stocks, and the S&P 500's first losing week in three. Away from the tickers, the frontier kept moving: a contested but independently-checked math proof from a single prompt, a $188B private valuation, and a run of ambitious open-weight models. The through-line is a widening gap between what public investors will pay for AI today and how fast the underlying capability and venture money are still moving.
The AI trade wobbles
The compute rally met its first real resistance. The Nasdaq fell 1.4% on Friday and the S&P 500 slid about 1% to close its first losing week in three, as investors began questioning whether AI valuations and the demand for memory and compute underpinning them can hold. Worries about "unsustainable" HBM and memory demand spread from NVIDIA to the Asian supply chain, and the sell-off dragged markets down as a roughly $1.8 trillion rally in Asian chipmakers — TSMC, SK Hynix, Samsung — started to unwind, with some large funds reported to be trimming exposure. It is only one week, but it is the first time in months the market has openly doubted the trade rather than chased it.
A machine reaches for a new theorem
While equities fell, the capability news pointed the other way. OpenAI's GPT-5.6 reportedly produced, from a single prompt, a novel proof cracking a convex-optimization problem open since the 1990s — a matching Ω(d²) lower bound arrived at after about 148 minutes of reasoning. Two independent groups confirmed the key lemma, one using CVXPY and one a Julia SDP solver, and the argument compiles in Lean 4. The result topped Hacker News and, if it stands, marks a step from AI that assists research toward AI that does the load-bearing part of it. The caveat matters: mathematicians are still debating whether the proof is genuinely new or a clever recombination of known techniques, so this is a reported and contested milestone, not a settled one.
Private money keeps marking up
The public pullback did not reach private valuations. Databricks is raising at a $188 billion valuation, a Coatue-led round of roughly $3 billion that lifts it 40% from the $134 billion it was worth in February, with the term sheet signed and the round expected to close later this summer. The company confirmed the strategic round, and the timing is the point: private AI-infrastructure investors kept paying up in the same week their public counterparts headed for the exits. When two markets price the same thesis in opposite directions, one of them is early.
Murati bets on open, not biggest
Mira Murati's Thinking Machines shipped its first model, and it is a statement of strategy more than of raw power. Inkling is a 975-billion-parameter mixture-of-experts with about 41 billion active parameters, released under an Apache 2.0 open-weights licence, natively multimodal across text, image, audio and video, trained on roughly 45 trillion tokens, and available on Hugging Face alongside the lab's "Tinker" fine-tuning platform. The startup, valued at around $12 billion, openly says the model "isn't the strongest available" — a deliberate wager that the future belongs to models teams can own and customise rather than to a single smartest general system.
China's open-weight cadence
The open-weight wave is not only Western. At the World Artificial Intelligence Conference, Alibaba previewed Qwen 3.8-Max, a 2.4-trillion-parameter open-weight multimodal flagship it claims is "second only to Anthropic's Fable 5" and ahead of every Western open model, with weights promised soon. It arrives just days after Moonshot's Kimi K3, and the reported ranking rests on Alibaba's own numbers — there is no model card, licence or active-parameter count yet. Treat the benchmark boast with caution; the harder-to-dismiss signal is the sheer pace at which frontier-scale open weights are now being released.
Capital moved two ways this week — down in public AI and chip stocks, up in private AI infrastructure — and that divergence is the thing to watch, because it cannot last indefinitely. The near-term catalysts: memory and HBM pricing plus hyperscaler capex commentary into the coming earnings, which will show whether the sell-off is a pause or a genuine demand turn; the mathematics community's verdict on the GPT-5.6 proof; and the actual release of Qwen 3.8-Max, with the model card, licence and independent benchmarks needed to test its claims. Signal, not advice; no live prices.