DeepSeek Raises $7.4B in Its First Outside Round — and Doubles Down on Open Weights
China's open-weight standard-bearer is about to be durably capitalised. DeepSeek is set to raise roughly 50 billion yuan ($7.4 billion) in its first-ever external funding round, valuing the lab at between $52 billion and $59 billion — one of the largest startup rounds in Chinese history (Bloomberg, SCMP). The cap table is its own signal: founder Liang Wenfeng is contributing about 40% himself (~20 billion yuan), with Tencent (~10 billion yuan) and battery giant CATL (~5 billion yuan) the largest external cheques, and the state-backed National Artificial Intelligence Industry Investment Fund taking part — a pointed endorsement from Beijing (Tech Startups, American Bazaar).
The strategy attached to the money is what matters for European buyers. DeepSeek's management has told investors it will prioritise frontier research over near-term commercialisation, with Liang reiterating a commitment to keep shipping open-source models while pursuing AGI (Bloomberg). For months this brief has tracked an open-weight cohort holding within a point or two of the proprietary frontier at a fraction of the hosted cost; the lab most associated with that trade just secured the balance sheet to keep weights free rather than fence them. It lands in the same week Brussels named an open-source software strategy a pillar of its Tech Sovereignty Package — and the economic case for open weights, not the ideological one, is precisely what makes a curated-OSS-plus-router platform deployable on infrastructure a European team controls.
Computex Turns to Physical AI — and Disaggregated Inference Steps Forward
With the show floor running through 5 June, Computex 2026's through-line is physical AI: humanoid robots, factory automation, and the supply strain of building for both (TechXplore, PR Newswire). NVIDIA leaned in with the Isaac GR00T Reference Humanoid Robot, an open reference design meant to lower the barrier to humanoid research, alongside Alpamayo 2 Super, a 32-billion-parameter open reasoning VLA model for robotaxis (TechXplore, BGR).
The quieter but more consequential thread for inference operators came from the non-NVIDIA corner. Intel unveiled new rack-scale AI infrastructure for inference and agentic workloads built on Xeon processors paired with SambaNova SN-50 RDUs, and a new entrant, Vector Core Compute — formed by Vista Equity Partners and Cambium Capital — demonstrated fully disaggregated inference running across Xeon CPUs, SambaNova RDUs, and NVIDIA Blackwell GPUs in one stack (Intel Newsroom). Disaggregation — separating prefill, decode, and KV-cache duties across heterogeneous silicon — is the architecture that turns inference from a single-accelerator bet into a tunable system, and seeing it shipped commercially is a useful signal for anyone whose cost structure lives or dies on tokens-per-watt rather than on which lab trained the model.
SpaceX Opens Its Roadshow at a $1.77T Valuation Markets Are Already Questioning
The largest IPO ever began marketing today. SpaceX opened its roadshow on 4 June at a fixed $135 per share, selling 555.6 million shares for a ~$75 billion raise at a $1.77 trillion valuation, with pricing set for 11 June and a Nasdaq debut on 12 June under SPCX (CNBC, TradingKey). At that level the company would rank among the seven largest in the US — above Tesla — and the offering is more than triple the size of the previous record US listing.
The valuation is contested out of the gate: Morningstar pegs fair value near $780 billion, less than half the IPO target, and advises long-term investors to wait for a more attractive post-listing entry (TechTimes). The AI relevance is the line buried in the prospectus: the Anthropic Colossus compute commitment — the multi-year, ~$45 billion deal underpinning Claude's most capital-intensive substrate — is a headline AI-revenue contributor for SpaceX's infrastructure arm. The first AI-era infrastructure IPO is being priced, in part, on a frontier lab's trajectory, which is a reminder of how tightly the compute supply chain and the model economics are now braided together.
Quick Hits
- CADA's tiered access draws the US-protectionism line — In the wake of the Tech Sovereignty Package, the live fight is over how much of Europe's public-sector cloud spend US giants keep. The Cloud and AI Development Act is expected to grant European providers preferential treatment for sensitive government workloads while letting US firms compete for less-classified projects; AWS, Microsoft, and Google — together ~70% of EU cloud — are already positioning with sovereign offerings (Microsoft's Bleu and Delos, AWS European Sovereign Cloud, Google air-gapped), while US voices frame the rules as disguised protectionism (Cryptobriefing).
- WWDC is T-4 — Apple's 8 June keynote is expected to centre on a chatbot-style Siri running on a custom 1.2-trillion-parameter Gemini model, reportedly for ~$1B/year, with the heavy lifting potentially on Apple's own private servers to wall user data off from Google (Macworld).
- Open-weight cost counter holds; MiniMax M3 weights imminent — The open cohort (DeepSeek V4, Qwen3.7 Max, GLM-5, Mistral Large 3) stays within a point or two of the Western frontier on agentic coding at a fraction of hosted cost; MiniMax M3's full weights and technical report are due around 11 June, the next reproducibility test (llm-stats).
- Mistral explores its own silicon — France's Mistral says it may design its own chips as it builds out inference data centres, and is guiding toward ~€1B in 2026 revenue (from ~€200M), the clearest sign yet of a European lab pushing for vertical control of its stack (Techzine, CNBC).
Editor's note: with the Tech Sovereignty Package now published, this brief is in the "reaction and reading" phase — so the one regulatory item (CADA's tiered-access fight) sits in quick hits rather than as a main, per the standing guidance to lead with Models, Enterprise, or Open Source when those buckets have credible stories. Today they did.