← Back to feed

The contract manufacturers that assemble AI servers for the world’s largest data centre operators have notified customers of price increases exceeding 15% in many cases, Bloomberg reported. The affected systems span Nvidia’s Vera Rubin and Grace Blackwell lines — the two hardware generations driving AI infrastructure buildout in 2026 and 2027. Microsoft, Google, and Oracle are among the operators receiving the new pricing. The increases apply to systems shipping in early 2027.

Nvidia did not respond to Bloomberg’s requests for comment.

The Memory Problem

The root cause is high-bandwidth memory, not chip shortage in the traditional sense. AI demand for HBM has outpaced supply from the industry’s three major memory manufacturers, pushing up prices across the board. Unlike earlier GPU shortages driven by chip fabrication constraints, this is a memory supply problem that affects the full server stack regardless of which processor sits at the centre.

That distinction matters because it also limits the insulation available to operators with in-house chip programmes. Alphabet designs its own tensor processing units and offers them through Google Cloud, reducing dependency on Nvidia directly. But Bloomberg noted that even companies running proprietary silicon still source HBM from the same three suppliers. Building your own chip does not exempt you from the memory market. The 15%+ increase lands on everyone’s stack.

The Capex Math

Alphabet raised its 2026 capex guidance in July to $195-205 billion, up from a prior $180-190 billion range. On that call, CFO Anat Ashkenazi said roughly 60% of second-quarter technical infrastructure investment went to servers — with the balance split across data centres and networking. At the $200 billion midpoint, that puts server purchases near $100 billion for 2026 alone.

That budget is probably committed. The price increases apply to 2027 shipments. What is not sized yet is 2027 capex, and Ashkenazi said explicitly on the July call that the company “continues to expect capex to increase significantly in 2027” without providing a figure.

For Alphabet, Google’s own server fleet buys some buffer. For Microsoft, which has no equivalent in-house silicon programme at scale and has committed to significant OpenAI infrastructure expansion, the exposure is more direct. For Oracle, which has been rapidly expanding its cloud infrastructure capacity as the third hyperscaler in the AI stack, the timing compounds an already aggressive build schedule.

What 15% Means

The increase varies by chip generation and memory configuration, so the 15%+ figure is a floor rather than a fixed number. At Alphabet’s scale — $100 billion in server spending at the midpoint of current guidance — even a 10% blended increase in 2027 server costs represents roughly $10 billion in additional spend before any increase in total capex.

The numbers get larger when Microsoft, Oracle, and other hyperscalers are added. The major cloud providers combined are on track to spend over $800 billion on AI infrastructure in 2026 according to earlier TrendForce projections. A 15%+ uplift on the server component of that, applied to 2027, reshapes the economics of the next generation of AI infrastructure significantly.

Whether that cost flows through to model pricing, cloud service rates, or absorbs into margins remains to be seen. What is clear is that the companies building out AI capability at scale are about to pay materially more for the hardware required to do it.