Rob FalconerThe latest in AI

Topic · Hardware & Industry

AI hardware & industry

Behind every model launch is a race for chips, power and data centres. This page tracks the infrastructure and business developments behind the headlines, using company announcements and reporting I can link to.

Last updated:

NVIDIA: an AI data centre in a shipping container

21 Oct 2026At GTC in Berlin, NVIDIA announced the DSX Pod AI Factory: seven Vera Rubin racks packed into a shipping container weighing about 28,000 kg, which can be delivered by lorry. NVIDIA says it can cut the time for a new AI data centre to start serving from about six months to 14 days. CNBC reports that deliveries are expected from early 2027 and that analysts estimate the cost at more than $7 million per rack.

7 racksVera Rubin, per container
~28,000 kgcontainer weight
14 daysNVIDIA's claimed time to serving
Early 2027expected deliveries

Google Cloud and TPUs

A5X on Vera Rubin NVL72

21 Oct 2026Google Cloud's A5X instances, built on NVIDIA Vera Rubin NVL72 systems, are due to become generally available next month, with expansion to Europe planned. NVIDIA says Google plans to deploy up to 960,000 Rubin GPUs. The first named customer is London-based Ineffable Intelligence, founded by David Silver.

OpenAI's models arrive on Google Cloud

21 Oct 2026Google Cloud now offers OpenAI's GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna in private preview, billed through Google Cloud and counting towards customers' spending commitments. Google says it is the only hyperscaler offering models from all three leading closed-model labs. For buyers, it means one more route to the same models, and for Google, a reason to keep big AI spend on its cloud.

TPU 8t and TPU 8i

22 Apr 2026Google's eighth-generation TPUs were announced in April for general availability "later this year". A TPU 8t superpod links 9,600 chips for 121 exaflops, according to Google.

Running frontier-scale models locally

At the other end of the scale, Meta says its open-weight Muse Spark Open (816B total parameters, 42B active) runs on two Apple Mac Studios with M5 Ultra chips. That is a notable change: a model of this size can run on equipment that a small business could buy.

15 Oct 2026NVIDIA is pushing the same way on PCs. Its Nemotron 3.5 Bolt 74B-A9B model fits in 24 GB of memory using native 2-bit quantisation, aimed at RTX graphics cards and DGX Spark, and a 4 GB Bolt E4B is due on 23 October. The weights, data and training recipes are open.

China: DeepInfer and Huawei Ascend

Oct 2026DeepSeek has published DeepInfer on GitHub, an inference engine for its DeepSeek-V4.1-Flash model. It supports NVIDIA's Blackwell-generation (SM100) GPUs and Huawei's Ascend 950 chips. The public release covers offline inference only. DeepSeek says its KV cache system, disaggregated serving and load balancing are not included. Support for a Chinese accelerator alongside NVIDIA is the notable part, in the same week that the US ordered federal civilian agencies to remove DeepSeek (see the policy page).

Europe's build-out

Mistral says Large 4 was trained on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centres, funded by a €3 billion Series D. Taken together with Google's planned A5X expansion to Europe and NVIDIA choosing Berlin for GTC, it shows the AI infrastructure race reaching Europe. In the UK, a London start-up is the first named A5X customer.

The economics: cheaper tokens, huge user numbers

  • Prices: OpenAI says GPT-6 Sol and Luna are 50% cheaper than GPT-5.6 promotional pricing. Anthropic says Opus 5.5 costs about 40% less than Opus 5 on typical workloads, and Haiku 5.5 about 75% less than Haiku 4.5. Gemini 4 Argon launched at an introductory $2 / $10.
  • Scale: OpenAI reports 1.2 billion weekly users and more than 3 billion images created a week in ChatGPT.
  • The cost of safety: OpenAI says safety monitoring of its models adds about 20% overhead, which shows up in computing budgets.
  • Speed for less: Inception prices its Mercury 3 diffusion model at $0.20 / $0.70 per million tokens, half that at launch, and claims it is 4 to 8 times faster than GPT-6 Luna and Claude Haiku 5.5.
  • Premium tiers: OpenAI's $500-a-month Pro 500 plan shows that heavy agent users are a market of their own.

What I left out

I could not confirm several widely repeated hardware claims from a primary source, so I have left them out for now. These include shipping dates for rival accelerators, some very large revenue and order-book figures, a reported large DeepSeek funding round and its reported daily token volume, and reports that OpenAI has changed its lobbying on chip exports. I will add them once I can link to a primary source.

Sources

  1. Nvidia announces Vera Rubin racks that can be shipped on a truck — CNBC (21 October 2026)
  2. Google Cloud and NVIDIA Vera Rubin infrastructure — NVIDIA Blog (21 October 2026)
  3. Google Cloud brings OpenAI models to customers — Google Cloud (21 October 2026)
  4. Eighth-generation TPUs for the agentic era — Google (22 April 2026)
  5. Nemotron 3.5 Super VL and Bolt — NVIDIA Blog (15 October 2026)
  6. DeepInfer — DeepSeek on GitHub (accessed 23 October 2026)
  7. Introducing Mercury 3 — Inception (19 October 2026)
  8. Muse Spark Open — Meta AI research blog (15 October 2026)
  9. Mistral Large 4 — Mistral AI (6 October 2026)
  10. Introducing GPT-6 Sol and Luna — OpenAI (22 September 2026)
  11. Claude Opus 5.5 — Anthropic (22 September 2026)
  12. Claude Haiku 5.5 — Anthropic (7 October 2026)
  13. Introducing Gemini 4 Argon (general availability) — Google (13 October 2026)
  14. DevDay 2026 recap — OpenAI (29 September 2026)
  15. Introducing ChatGPT Images 2.5 — OpenAI (8 September 2026)
  16. Pacing model development around cyber capabilities — OpenAI (18 August 2026)

Figures quoted from company announcements are the companies' own claims. Independent testing may give different results.