The dated market snapshot. Factual snapshot: August 2026. This document ages by design — it is refreshed each revision so the rubrics in Documents 2 and 3 do not have to be. Sources for every dated claim: Document 5, Evidence Base.
1. Models
Frontier and open weights. Capable open-weights models are cheap and common: Llama 4, Gemma 3, Qwen 3, DeepSeek V3/V4, Mistral. DeepSeek-V4 (April 2026 preview; reported 1.6T-parameter MoE, ~1M-token context, MIT-licensed weights) shipped with day-0 adaptation to Huawei Ascend, Cambricon, Hygon, and Moore Threads — the first top-tier model optimised for a fully non-NVIDIA stack rather than treating it as fallback. Reasoning-capable models (DeepSeek R1, EXAONE Deep, Sarvam-M, Qwen thinking mode) are now a separate marker of national capability.
The Fable 5 / Mythos 5 episode (June–July 2026). A US export-control directive suspended all foreign-national access to Anthropic’s frontier models on 12 June; the directive was lifted 30 June and Fable 5 restored globally 1 July, with Mythos 5 limited to approved US organisations. The 19-day outage is the framework’s flagship live Version Deadlock case (Framework §3.4) and drove a wave of European sovereign-AI initiatives: Austria’s proposal that the EU host Anthropic capacity, the UK’s Isambard-trained “Lumen Sovereign” frontier model, and accelerated German investment around Aleph Alpha.
National models. BgGPT 3.0 (Bulgaria, March 2026) on Gemma 3; Meltemi (Greece), Sarvam-M (India), TAIDE (Taiwan) at L2; EuroLLM, Teuken, Poro, Aya-Expanse, EXAONE at L3; Falcon/Jais (UAE), Fugaku-LLM (Japan), DeepSeek/Qwen-on-Ascend (China) at L4. Mistral remains Europe’s only top-10 global model firm (€11.7B valuation after the September 2025 ASML-led €1.7B Series C).
Reference costs. DeepSeek-V3’s reported training cost (~$5.6M rented compute, 2.788M H800 GPU-hours) is the floor for replicating a top model; ~$500M/year approximates a full national ecosystem.
2. Cloud and sovereign infrastructure
Certified sovereign tier. SecNumCloud 3.2 qualified providers: OVHcloud, Outscale, Scaleway, and S3NS (Thales + Google, qualified 17 December 2025 — the first major US-technology service to pass); Bleu (Capgemini + Orange + Microsoft) still in qualification.
EU-operated US subsidiaries. AWS European Sovereign Cloud (GA 15 January 2026, Brandenburg, €7.8B): EU staff, EU-citizen directors, but Amazon-owned — CLOUD Act exposure arguably remains; qualifies at the BSI C5 tier. Azure EU Data Boundary is a contractual promise; Microsoft’s own French Senate testimony (June 2025) conceded CLOUD Act protection cannot be guaranteed.
The EU policy layer. The Cloud and AI Development Act (proposed 3 June 2026, part of the Technological Sovereignty Package with Chips Act 2.0) would create a four-level Cloud Sovereignty Framework for public procurement and aims to triple EU data-centre capacity in 5–7 years; in trilogue. The Commission’s first sovereignty-criteria cloud procurement (€180M, April 2026) went to four European groups. EUCS remains stalled with its sovereignty tier removed.
Compute programmes. 19 EuroHPC AI Factories across 16 states (~€2.6–2.7B committed) plus 13 Antennas. The AI Gigafactories call opened 30 July 2026: seven sites, ≥100,000 advanced chips each, €10B public targeting ≥€20B private; bids close 12 November 2026, awards early 2027, operations ~18 months after. Ten member states have signalled hosting interest. Outside the EU: Stargate UAE (1 GW, first 200 MW due Q3 2026) inside the 5 GW UAE–US AI Campus.
Constraints. Roughly 95% of commercial AI compute is US- or Chinese-owned. European data centres cluster in FLAP-D plus Madrid and Milan; grid capacity is a live constraint (Irish data centres exceeded 20% of national electricity in 2024).
3. Hardware
NVIDIA. Blackwell GB300 in full production; Vera Rubin platform launched at CES 2026 (NVL72 rack: 3.6 EFLOPS inference, ~$8.8M/rack, partner availability H2 2026). Vendor figures, not independently tested.
AMD. MI300X/MI325X in production at scale (Azure, Meta, El Capitan); MI355X clusters at Oracle up to 131,072 chips; ~$100B Meta–AMD deal (late 2025); MI400/Helios expected H2 2026. ROCm 7.x makes AMD a believable second source.
Hyperscaler ASICs — the real threat to NVIDIA’s grip. Google TPU v7 Ironwood (GA November 2025; a reported 1M-chip Anthropic order); AWS Trainium3 (3nm; 500k-chip Anthropic cluster); Microsoft Maia 200; Meta MTIA v4 (RISC-V, with Broadcom). TrendForce forecasts custom-ASIC shipments growing 44.6% in 2026 vs 16.1% for GPUs, with NVIDIA’s inference share possibly falling to 20–30% by 2028.
Chinese stack. Huawei Ascend 910C (~H100-class on paper; ~600k units 2026); Ascend 950PR shipping, 950DT due Q4 2026; Cambricon ~500k chips. With DeepSeek-V4’s day-0 Ascend adaptation, China’s top models no longer need CUDA.
European and Korean alternatives. SiPearl Rhea1 sampling (JUPITER); Rhea2 for Jules Verne; EU DARE RISC-V programme (~€240M); Tenstorrent (open RISC-V, Blackhole shipping); Rebellions ($400M at $2.3B pre-IPO); Cerebras ($10B+/750 MW OpenAI partnership, January 2026).
US export-control posture. Seven major turns 2022–2026, now deal-by-deal: the H20 15% and H200 25% revenue-share deals (2025); case-by-case licence review for H200-class (January 2026); the offshore-subsidiary rule (June 2026 — licensing requirements follow Chinese-headquartered firms abroad); H20 licences granted (July 2026). Uptake is another matter: as of mid-May 2026, no H200s had reportedly shipped to approved Chinese buyers.
4. Manufacturing and materials
Leading edge. TSMC N2 in mass production (Q4 2025; ~$25–27k/wafer; booked through 2028; Apple holds ~half of 2026–27 supply). Samsung 2nm ramping with yield struggles. Intel 18A ramping (Panther Lake shipped; AWS and Microsoft as customers) — but Intel cancelled Magdeburg and Wrocław (July 2025), forfeiting ~€10B in subsidies; ESMC Dresden (TSMC-backed, 28/22 + 16/12nm) proceeds. Chips Act 2.0 (proposed 3 June 2026) is the EU’s response: faster permitting (≤12 months), investment conditions, “Grand Challenges” including AI chips. The European Court of Auditors projects the EU at ~11.7% of the global chip value chain by 2030, versus the 20% target.
Packaging and memory — the real bottlenecks. CoWoS: ~70k wafers/month end-2025, targeting 130k end-2026, NVIDIA holding ~half. HBM4: SK Hynix, Micron, Samsung all booked for 2026; the main constraint on Rubin-class parts.
Critical minerals. China refines ~89% of rare earths and ~60% of germanium. The October 2025 controls were paused in November 2025 for 12 months — but enforcement through 2026 has been active: new restrictions on 10 US companies (June) and 14 EU firms (July), formalised violation-handling rules, and the first detentions of foreign nationals over alleged rare-earth smuggling. The IEA warns full enforcement could put $6.5T of downstream production at risk. The EU is standing up a joint critical-minerals purchasing and stockpiling centre. The Nexperia affair (Dutch state supervision from September 2025; Chinese counter-controls; the company operationally split by March 2026) shows the same weapon pointed both ways.
5. Software
CUDA and alternatives. NVIDIA holds ~75–90% of the accelerator market; CUDA beats ROCm by 10–30% on compute-bound work, less on memory-bound. TorchTPU (Google + Meta, December 2025) brings full PyTorch support to TPUs — the most direct assault on CUDA lock-in. PyTorch’s multi-party foundation governance (AMD, AWS, Google, Meta, Microsoft, NVIDIA on the board) makes single-vendor weaponisation harder.
Inference runtimes. vLLM: best TTFT, broad hardware support — the default. TensorRT-LLM: 10–30% more throughput on pure-NVIDIA deployments. SGLang: RadixAttention, up to ~6.4× throughput on shared-prefix/agent workloads. HuggingFace TGI: maintenance mode — avoid.
6. Regulatory timeline (as of August 2026)
| Instrument | Status |
|---|---|
| EU AI Act — bans, AI literacy | In force since 2 February 2025 |
| EU AI Act — GPAI obligations | In force 2 August 2025; enforceable (AI Office penalty powers, Article 50 transparency) since 2 August 2026 |
| EU AI Act — high-risk | Deferred by Digital Omnibus (in force 27 July 2026): Annex III → 2 December 2027; Annex I → 2 August 2028 |
| GPAI Code of Practice | Endorsed 1 August 2025; Meta declined; xAI signed safety chapter only |
| EU Data Act Chapter VII | Fully applicable since September 2025 |
| DORA | In force since 17 January 2025 |
| Council of Europe AI Convention (CETS 225) | In force 1 November 2025 |
| CADA + Chips Act 2.0 | Proposed 3 June 2026; in trilogue |
| New Delhi Declaration | Adopted 18–19 February 2026; 92 endorsements |
| US posture | Deal-by-deal export diplomacy; Fable/Mythos directive (12 June) lifted 30 June 2026 |
7. Human capital and talent reference figures
MacroPolo Global AI Talent Tracker 2.0 (March 2024, NeurIPS 2022 data): 75% of top AI researchers at US institutions did undergraduate work in the US or China; China produces 47% of top researchers by undergraduate origin; mobility is falling. Stanford AI Index 2026: the US holds ~75% of global GPU compute (up from 51%); China’s share fell to ~14%; the US produced 50 notable models in 2025 vs China’s 30; Switzerland (110.5) and Singapore (109.5) lead researchers per 100k population; the US–China top-model capability gap has all but closed; 80 of the 95 most notable 2025 models shipped without training code; average Foundation Model Transparency Index score fell to 40.