Editor's note: today's three main stories are Open Source/Models, Enterprise/Infrastructure, and one Regulation item (the EU's Google decisions). We lead with the model and the silicon; the regulation main earns its slot because it is the most concrete EU action on AI-assistant competition this week and sits squarely on the through-line — this week the fight for control of the AI stack played out at three layers at once: the model, the chip, and the distribution gate.
Alibaba's Qwen3.8-Max Makes It Two Trillion-Parameter Open-Weight Claims in a Week
Three days after Moonshot AI unveiled the 2.8-trillion-parameter Kimi K3, Alibaba used the closing days of the World AI Conference in Shanghai to preview Qwen3.8-Max, a 2.4-trillion-parameter multimodal model it says is “second only to Fable 5.” It would be Alibaba's first trillion-parameter-plus model to handle images, video, and documents, and — like K3 — it is destined for open weights. On paper, that is a remarkable fortnight: the two largest publicly announced open-weight models to date, both from Chinese labs, both landing inside a week.
The asterisk is the same one we keep flagging. Qwen3.8-Max shipped with no model card, no independent benchmarks, no license, and no standard per-token price — the “second only to Fable 5” ranking is Alibaba's own internal read. Today it is a preview reachable only through Alibaba's Token Plan and Qoder at a tenth of standard pricing; the open weights are “coming soon” with no date. A model becomes catalogue-able for a regulated EU buyer when the checkpoint is downloadable, the license is inspectable, and a neutral board has scored it — not when the press release goes out. Until then, both K3 (weights due 27 July) and Qwen3.8-Max are announcements, not artifacts.
Etched Nears a $20 Billion Valuation Betting the Inference Layer Leaves the GPU Behind
The layer beneath the model is drawing its own capital. Transformer-chip startup Etched is in talks for two simultaneous rounds — one led by Sequoia at roughly $10 billion, another led by existing investor Jane Street at about $20 billion, per the Wall Street Journal. The higher figure would quadruple the $5 billion valuation Etched carried when it exited stealth barely three weeks ago. Neither deal had closed as of mid-July, and terms may move, but the direction of travel is clear: investors are pricing a serious bet against general-purpose GPUs for inference.
Etched's pitch is narrow by design. Its Sohu chip hard-codes the transformer architecture into silicon, and the company's own launch figures — not independently benchmarked — claim an eight-chip server processes over 500,000 tokens per second on Llama 70B against roughly 23,000 for an equivalent eight-GPU H100 box. The trade-off is inflexibility — a chip that only runs transformers is worthless the day the dominant architecture changes. For European operators the relevant read is not the throughput number but the structural one: the cost and control of inference are increasingly decided below the model layer, and a hardware market that is less monolithic than “buy NVIDIA” is a precondition for anyone building infrastructure they actually own.
Brussels Tells Google to Open Android to Rival AI Assistants
The one regulation main this week is also the most concrete EU action on AI competition in months. Under the Digital Markets Act, the European Commission issued two binding decisions on 16 July requiring Google to open eleven Android feature groups to rival AI assistants — the system-level hooks it had reserved for Gemini. Competing assistants will be able to be summoned by voice, act inside apps on a user's behalf, and reach the functionality that made “Hey Google” a moat. A parallel decision requires Google to share anonymized Search data with competing search engines, explicitly including AI chatbots with search features.
Google objected on the same day, warning through policy chief Kent Walker that the data-sharing could weaken privacy and expose “private searches” to unfamiliar firms. The timelines are real but not immediate: the search-data pricing offer and access must be finalized by January 2027, and all eleven Android features must reach qualifying rivals in Android 18 by August 2027. The sovereign read is familiar from the other direction — where the AI Act governs how models behave, the DMA governs who controls the surfaces they run on. For once, the answer to “who sets the terms of access” is a European regulator rather than a US platform.
Quick Hits
- Gemini 3.5 Pro slips a third time — Google's flagship missed its reported 17 July target after DeepMind reportedly scrapped and rebuilt the base model over coding and hallucination shortfalls; prediction markets now lean late-July to early-August. Still internal-and-enterprise-preview only.
- EU AI Act GPAI enforcement goes live 2 August — in twelve days the Commission's supervision and penalty powers over general-purpose-model providers become applicable, with fines up to €15M or 3% of global turnover under Article 101 and the power to demand documentation, run evaluations, or restrict a model's EU market access.
- Kimi K3 weights due 27 July — the artifact test for Moonshot's “world's largest open model” arrives in six days; the checkpoint, license terms, and first neutral benchmark reads will decide whether the 2.8T claim is deployable or just headline-sized.
- Grok 4.5 lands #4 on the Intelligence Index — xAI's first post-SPCX flagship now has independent Artificial Analysis scores at Intelligence Index 54, behind Fable 5, GPT-5.5, and Opus 4.8 — a reminder that neutral scores, not launch tables, settle the leaderboard.