Editor’s note: today’s lead rotates off yesterday’s US-earnings print to a European story on a distinct axis — the compute supply chain. Two of the three main items sit in infrastructure because that is where the fresh, dated news landed in the last 24–48 hours; the through-line is that capability is converging and cheapening while the durable questions move down the stack to who makes the chips and where the weights run. Regulation is held to a single quick hit — the EU AI Act’s new enforcement powers are live, but there is still no named action.
Europe’s biggest-ever chip round bets on light, not HBM
UK photonic-computing startup OLIX raised a $312M Series B at a $3.3B valuation — reportedly the largest semiconductor venture round ever raised by a European company — with the UK government’s Sovereign AI Fund, Arm, and Hudson River Trading among the backers, alongside Netflix co-founder Reed Hastings and existing investors who all expanded their positions. The pitch is a direct attack on the supply chain that gates today’s AI: OLIX’s “Optical Tensor Processing Unit” holds models in fast on-chip SRAM and computes with light, removing high-bandwidth memory — the costly, supply-constrained part underpinning Nvidia’s top GPUs — from the architecture entirely.
For European buyers this is the sovereignty story at the layer that usually gets ignored. Data residency and model licensing get the headlines, but the hardest dependency to unwind is the accelerator supply chain, which runs through a handful of US and Taiwanese chokepoints. A credible European alternative-architecture bet, part-funded by a national sovereign-AI vehicle, is the kind of structural move that matters more than another model launch — with the obvious caveat that DX-1 silicon is not slated for initial customer access until the second half of 2027. The capability is a roadmap, not a product you can rack today.
AMD’s record quarter shows the second accelerator source is real
AMD reported Q2 revenue of $11.5B, up 50% year over year and a company record, beating consensus, with adjusted EPS of $1.66. The data-center segment more than doubled to $6.7B (up 107%) and now accounts for 58% of the company — a concrete signal that Nvidia is no longer the only serious supplier of AI-training and inference silicon at scale. AMD guided Q3 to $12.7–13.3B, roughly 41% growth, and said its Helios rack-scale system enters shipments toward the end of the quarter.
The customer list is the point: Helios racks — 72 MI455X GPUs, 31TB of unified HBM4, 2.9 exaflops of FP4 compute — are contracted to Microsoft, Meta, OpenAI, and Oracle, with Anthropic’s deal covering up to 2GW of MI450-series parts plus up to $5B in AMD equity. That is genuine supply diversification at the chip layer, and it should ease the capacity crunch that has throttled inference pricing. But note where the capacity goes: overwhelmingly to US hyperscalers and US-jurisdiction labs. A second silicon vendor widens the pipe; it does not, by itself, move the compute onto European soil — which is exactly the gap OLIX and the sovereign-cloud buildouts are trying to close.
The open-weight field keeps closing the gap you can actually own
While the hardware bill climbs, the capability that European enterprises can hold on their own infrastructure keeps improving. Alibaba’s Qwen3.8-Max — announced 3 August, a 2.4T-parameter sparse MoE (~95B active) with a 1M-token context and multimodal input — is billed as the first Max-class Qwen slated for open weights, with the checkpoints (plus a smaller Qwen3.8-27B) due on Hugging Face and ModelScope this week. Today it is API-only via DashScope, and the benchmark table is Alibaba’s own, not an independent reading, so treat “open-weight” as a pledge until the repo says otherwise.
The through-line ties back to the compute-supply story. Kimi K3 already leads open-weight models on public reasoning benchmarks, DeepSeek’s V4-Flash-0731 shipped self-hostable under MIT last week, and if Qwen3.8-Max’s weights land as promised, the largest openly licensed model yet reaches frontier-class scores. For a regulated European buyer the calculation is unchanged: a converging, cheapening frontier means the durable variables are portability, jurisdiction, and whether you can run the model where your data lives — and downloadable weights are only sovereign if you have the racks to serve them, which loops straight back to who supplies the silicon.
Quick hits
- Regulation: The EU AI Act’s GPAI enforcement powers have been live since 2 August — the AI Office can compel documentation, run its own evaluations, and order mitigation or market withdrawal, with fines up to €15M or 3% of global turnover (Art. 101, the GPAI cap — not the €35M/7% Art. 99 prohibited-practices cap). Coverage this week confirms “technical compliance dialogues” as the opening tool; there is still no named enforcement action.
- Governance: Anthropic named former California Supreme Court justice Mariano-Florentino (Tino) Cuéllar as Chief Global Affairs Officer (4 August) — labs are staffing up policy capacity just as the enforcement window opens.
- Open-source watch: Qwen3.8-Max and Qwen3.8-27B weights are promised on Hugging Face and ModelScope this week; verify the artifact, not the announcement, before treating either as self-hostable.
- Next data point: Nvidia reports Q2 FY27 on 26 August — the read-through from AMD’s data-center surge to the market leader will re-open the capex-versus-ROI debate that has run through this week’s coverage.
- GLM-5.5 watch (recurring): Zhipu’s ~1T open-weight model is still widely expected in August on the two-month cadence, but remains unconfirmed by Zhipu — no card, endpoint, or date. Rumour until a repo says otherwise.