AI infrastructure news, week of 2 November 2026: Microsoft rethinks rented compute, NVIDIA's reported 51% rule, Marvell buys Lumentum

Microsoft drops Oracle and Nscale capacity, NVIDIA's reported China rule, Marvell buys Lumentum, Anthropic buys generators. What it means in Europe.

Green-clad Microsoft data centre building with cooling pipework behind a black fence in Middenmeer, the Netherlands
Photo: Microsoft Data Center Middenmeer, the Netherlands - 1 (cropped) by Hay Kranen, Wikimedia Commons, CC BY 4.0

Five stories from the week of 2 to 8 November 2026 that matter if you rent or buy GPU servers in Europe. Several rest on reports citing unnamed people, and we say so. Sources and dates are at the end.

This week in short

  • Microsoft reportedly declined an option to more than double its Nscale capacity in Texas, and is winding down a 2023 deal to rent AI capacity from Oracle.
  • NVIDIA has reportedly told cloud partners that at least 51% of their services must go to non-Chinese companies. Reuters could not immediately verify the report.
  • Marvell agreed to buy the laser maker Lumentum in an all-stock deal with an enterprise value of $115 billion, expected to close in the first half of 2028.
  • Anthropic reportedly paid $3.4 billion for 1.7 GW of gas-fired generators, a sign of how far AI companies now go to secure power.
  • Sam Altman met Samsung and SK leaders in South Korea, days after SK hynix said long-term agreements already cover more than half of its 2027 and 2028 capacity.

Microsoft steps back from rented capacity at Nscale and Oracle

On 5 November the Financial Times reported, and Reuters followed citing a person familiar with the matter, that Microsoft will not exercise an option to substantially expand its deal with Nscale at a Texas data centre. The extension would have more than doubled the rented computing power and could have been worth tens of billions of dollars to Nscale. The existing contract stays in place. The source said Microsoft decided months ago, before Nscale's IPO filing in September, that Nscale's roughly $103 billion total contract value is unaffected, and that Nscale is talking to several major customers about the remaining capacity.

A day later Bloomberg reported that Microsoft declined to renew a 2023 deal to rent AI capacity from Oracle, originally used for AI features in Bing, and has been switching it off in recent months. People familiar with the matter said the older NVIDIA chips no longer matched what Microsoft needed. Microsoft declined to comment. Oracle answered that demand for its AI infrastructure remains strong, including for older chips. On its September earnings call, Oracle's co-CEO Clay Magouyrk said all GPU capacity up for renewal in the quarter had been renewed or resold above the previous price, by about 20%, and most of those chips were four or more years old.

Our reading, not a forecast: this is not evidence of falling demand, because both stories are about what one very large customer wants and Oracle's numbers point the other way. But a neocloud's order book can change quickly when one customer dominates it. According to Bloomberg, as relayed by The Next Web, Microsoft has committed more than $80 billion to CoreWeave, Nscale and Nebius and is reassessing those contracts. If you depend on a neocloud, it is reasonable to ask how much of its capacity plan rests on a single tenant.

NVIDIA's reported 51% rule for cloud partners

The Financial Times reported on 3 November, citing people with knowledge of the matter, that NVIDIA has told data centre and cloud partners that at least 51% of their services must be sold to non-Chinese companies. According to the report, the message came verbally from sales and regional teams with no written notice. It applies to customers worldwide and to future contracts, not existing leases. Reuters could not immediately verify it. NVIDIA said in a statement that US policy supports the use of US infrastructure around the world, without addressing the 51% figure in the coverage we read.

The reports we read do not explain how the share would be measured or audited. Our assumption: European providers of NVIDIA capacity will ask more questions about who the end customer is, and buyers should have that answer ready.

Marvell agrees to buy Lumentum for $115 billion

Marvell and Lumentum announced a definitive all-stock agreement on 2 November. Each Lumentum share converts into 4.5 Marvell shares, about $1,132 per share and an implied enterprise value of $115 billion. The deal needs shareholder and regulatory approval and is expected to close in the first half of 2028. The release does not mention NVIDIA, which announced a $2 billion investment in Lumentum on 2 March, with a multibillion purchase commitment and future capacity rights for its lasers, and a $2 billion investment in Marvell on 31 March.

This changes nothing for buyers in the near term, given the 2028 close. It matters mainly for large clusters, where optics link thousands of GPUs. A single 8-GPU server is much less exposed. The wider signal is that lasers have joined memory and packaging on the list of parts that NVIDIA pays to secure in advance.

Anthropic buys 1.7 GW of gas-fired generators

The Financial Times reported on 8 November, citing a person with knowledge of the deal, that Anthropic agreed in late September to pay $3.4 billion for 1.7 GW of heavy-duty gas-fired generators from OnePWR Solutions of San Antonio, Texas. By our arithmetic that is about $2,000 per kW, including the engineering and construction capability to install them. OnePWR had ordered the machines from Caterpillar about two years earlier but needed money to complete the purchase. The FT called it the largest known single purchase of power equipment by an AI company. It is not known where the generators will go, and both companies declined to comment.

For Europe the lesson is the one from Microsoft's exit from Nscale's Loughton site last month: power decides where capacity appears and when. Our sovereign AI overview covers the EU's compute plans. When a provider promises capacity in a future building, ask to see the energised grid connection or the generator contract, not just the plan.

Altman in Korea as memory makers commit their output

Sam Altman met about 50 senior Samsung executives in Yongin on 31 October and SK Group chairman Chey Tae-won, eeNews Europe reported. Samsung announced no new memory contract. On 28 October SK hynix said long-term agreements with 14 major customers already cover more than half of planned 2027 and 2028 capacity, according to TrendForce. Korean press quoted by TrendForce described DRAM and HBM output as effectively sold out through 2029 once demand still under negotiation is counted. Its average DRAM selling price rose by a high-20s percentage in the third quarter.

Memory is a large part of a server's price, not only of the GPU. In our view, buyers should expect short validity on server quotes and should check exactly which DIMMs a quote includes. Our overview of RTX PRO 6000 server prices in Europe shows where the market stood this autumn.

Questions

Will Microsoft's decisions make GPU rental cheaper? We do not know. Some older capacity may return to the market, but Oracle reports reselling exactly such capacity at higher prices. No source we found gives a volume.

Where does NEOXIS fit? We run one 8 × RTX PRO 6000 server in a data centre near Sofia, from $1.80 per GPU-hour or about $10,510 a month for the whole machine. It is rented into early 2027 and we accept reservations for the period after that. Details are on the pricing page.

Sources

All pages were checked on 9 November 2026. We read the Financial Times and Bloomberg reports through the outlets named below, not in the original.

SHARE