AI infrastructure news, week of 19 October 2026: NVIDIA buys capacity, China orders RTX PRO cards, Microsoft leaves an Essex site
NVIDIA's supplier deals, $35bn of Chinese RTX PRO 5500 orders, Microsoft leaves an Essex site, Nebius borrows $3.5bn. What it means in Europe.

Four stories from the week of 19 to 25 October 2026 that matter if you rent or buy GPU servers in Europe. Each one is checked against the company's own release or the original Reuters report; the sources and dates are at the end.
This week in short
- NVIDIA signed capacity deals with six suppliers of substrates, packaging, testing and optics. Its own words: "Capacity is the constraint."
- ByteDance, Alibaba and Tencent ordered more than 3 million RTX PRO 5500 cards, worth over $35 billion, according to Reuters. NVIDIA reportedly warned it may fill only a small part.
- Microsoft walked away from Nscale's planned Loughton data centre in Essex after grid officials said power would not come before the 2030s.
- Nebius took a $3.5 billion, 364-day secured loan from Citi to pay for data centres and GPUs.
NVIDIA signs capacity deals with six suppliers
On 21 October, at GTC Berlin, NVIDIA announced plans with Amkor, IBIDEN, King Yuan Electronics (KYEC), Samsung Electro-Mechanics, SPIL and Unimicron to expand manufacturing of the parts around its chips. The substrate agreements with IBIDEN, Samsung Electro-Mechanics and Unimicron should together lift substrate capacity 2.5 times over three years. Amkor targets more than 300% more interposer wafer capacity in South Korea by the end of 2028. NVIDIA is also paying KYEC a $1.1 billion capacity reservation advance for new US testing facilities, with high-volume production targeted for April 2027, and it expects its total test capacity to double from 2026 to 2027.
The timeline is the useful part. Unimicron's new substrate capacity starts in the second quarter of 2027, IBIDEN's in the first quarter of 2028, and Amkor's packaging expansion runs to the end of 2028. In other words, NVIDIA itself expects the bottleneck in the steps after the silicon wafer to last well into 2027. For a team planning a GPU purchase in Europe, our reading is simple: do not plan around a sudden glut of new data centre GPUs next year. Lead times and prices for complete servers are more likely to stay firm than to fall. Our overview of RTX PRO 6000 server prices in Europe shows where the market stood this autumn.
China orders more than 3 million RTX PRO 5500 cards
Reuters reported on 20 October, citing anonymous sources, that ByteDance, Alibaba and Tencent have ordered more than 3 million NVIDIA RTX PRO 5500 cards at about $11,400 each, a total list value above $35 billion. Orders from all Chinese customers exceed 6 million cards, two sources said. The buyers want the cards mainly for AI inference, because US export rules block the B300 and put the H200 under case-by-case licensing.
NVIDIA told customers it cannot guarantee delivery. One source said partners may receive only 5% to 10% of what they asked for; another said NVIDIA could make no more than 3 million RTX PRO 5500 cards worldwide in 2027. Samples are due in December and commercial deliveries from the first quarter of 2027.
The RTX PRO 5500 is a workstation card from the same Blackwell RTX PRO family as the RTX PRO 6000. If the Reuters figures hold, demand for professional Blackwell cards outside the big training clusters is far larger than planned output. Our assumption, not a forecast: European buyers of RTX PRO cards and servers should expect tight allocation through 2027, and should treat a firm delivery date as part of the price.
Microsoft leaves Nscale's Loughton site over grid power
On 18 October The Telegraph reported, and Reuters followed, that Microsoft had ended its agreement to anchor Nscale's planned data centre in Loughton, Essex. The site was announced during the US president's state visit to the UK in September 2025 and was meant to hold 23,040 NVIDIA GB300 GPUs, about 50 MW of AI computing for Microsoft's Azure services in Britain. According to the Telegraph report, grid officials told Nscale they did not expect to supply power until the 2030s.
Microsoft said the decision "relates to a single site" and that it remains committed to its partnership with Nscale. Nscale said the relationship "remains strong and long-term" and that it is looking at alternative power sources and other UK sites. Nscale filed for a US stock market listing in September, with Microsoft and Anthropic named as expected major customers.
The lesson for European buyers is about power, not about Microsoft. A GPU contract is only as good as the electricity behind it, and grid connections for new sites in parts of Europe now take years. When a provider promises capacity in a building that is not yet finished, ask whether the grid connection is signed and energised, not only planned. Capacity that runs today in an existing data centre carries a different risk. We wrote more about this and about the EU's own compute plans in sovereign AI in Europe.
Nebius borrows $3.5 billion from Citi
Nebius, the Amsterdam-headquartered GPU cloud, announced on 22 October a $3.5 billion senior secured facility arranged by Citi. It runs for 364 days, starts at Term SOFR plus 2.75% a year with a ceiling of 3.50%, and the money goes to data centre development, GPU purchases and general corporate purposes.
Short secured loans like this are how the large GPU clouds bridge the gap between buying hardware and getting paid by customers. It shows that the money for new European GPU capacity is still available. It also shows the cost: buildout is financed with debt at a few percentage points over the base rate, and that cost ends up in rental prices. If you are comparing a neocloud with a dedicated server, the useful question is less "who is bigger" and more "what exactly do I get, for how long, and at what fixed price".
What this means for a European GPU renter
All four stories point the same way: bottlenecks NVIDIA expects to last into 2028, demand far above output, and new capacity that depends on power and debt. None of that suggests cheaper or easier GPUs in the next twelve months.
For a team that needs GPUs for a defined project, our view is to secure capacity that already exists and runs, at a price agreed in advance. NEOXIS runs 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.
Questions
Does the RTX PRO 5500 order affect RTX PRO 6000 supply? We do not know. Reuters reports on RTX PRO 5500 orders and NVIDIA's limits for that card only. Both are Blackwell products made at TSMC, so heavy demand for one can plausibly affect allocation of the other, but no source we found says so.
Will GPU server prices fall in 2027? NVIDIA's own timelines put the new substrate, packaging and testing capacity in 2027 and 2028. That makes a quick price drop in the next year unlikely, in our view, though prices of older cards can still move.
Sources
All pages were checked on 26 October 2026.
- NVIDIA Newsroom: NVIDIA and Supply Chain Partners Expand Capacity to Accelerate AI Infrastructure Buildout, 21 October 2026.
- Reuters, via BNN Bloomberg: ByteDance, Alibaba, Tencent place over US$35 billion worth of Nvidia chip orders, 20 October 2026.
- Reuters, via The Star: Microsoft pulls out of UK data center deal with Nscale, Telegraph reports, 18 October 2026, based on a report in The Telegraph.
- Nebius Group N.V.: Form 6-K, Exhibit 99.1, US SEC, 22 October 2026.
- NEOXIS: pricing and server specifications.


