FR PT
Accueil › Non classé › Compute and power: what supplier announcements change for AI capacity

Compute and power: what supplier announcements change for AI capacity

TL;DR On the compute side, Qualcomm announced two rack-scale inference solutions in October 2025, AI200 (2026) and AI250 (2027): 768 GB of memory per card, direct liquid cooling, 160 kW per rack. On the power side, Siemens is investing more than $200 million in two US plants building equipment for AI data centers, and ABB is supplying VoltaGrid with 62 synchronous condensers. Compute gains a supplier; power gains manufacturing capacity. None of these announcements publishes manufacturing lead times or prices.

An inference cluster depends on two separate supply chains. One delivers the accelerators and the racks that hold them. The other delivers everything that brings in, distributes and stabilizes electricity between the grid connection and those racks.

In September we covered each chain on its own on LinkedIn: Qualcomm’s entry into inference chips on 15 September, and the Siemens and ABB investments in electrical equipment on 25 September. Read side by side, from the manufacturers’ own releases, the two moves follow different dynamics, and each leaves different questions open.

That difference matters for an 18 to 36 month capacity plan. The two layers are not moving the same way, and the available information does not support the same conclusions.

Compute: one more supplier for inference

What Qualcomm announced

In October 2025, Qualcomm introduced two data center inference solutions sold at rack scale: AI200 and AI250.

AI200 carries 768 GB of LPDDR memory per card. Qualcomm says AI250 will deliver more than 10 times higher effective memory bandwidth. Both use direct liquid cooling, with power consumption of 160 kW per rack.

The announced timeline puts AI200 in 2026 and AI250 in 2027, on an annual product cadence. A first customer was named: HUMAIN, which plans to deploy 200 MW of these solutions in Saudi Arabia starting in 2026.

An inference-first positioning

Qualcomm describes these products as purpose-built for inference. Its pitch rests on three criteria: total cost of ownership, power efficiency and independence from suppliers.

Nvidia and AMD remain the established suppliers in this market. Qualcomm’s arrival adds a third name to the rack-scale inference options, with first delivery announced for 2026.

What still has to be measured

The specifications above are the ones Qualcomm has announced. In the sources we reviewed, we found no independent measurements of performance or power consumption in production.

Two things are worth tracking: the first deployment results, including HUMAIN’s, and whether AI200 ships on its 2026 schedule. AI250 is still an announcement, targeted for 2027.

Power: established suppliers adding capacity

Siemens and its US plants

On 7 August 2026, Siemens announced more than $200 million of investment in its plants in Pendergrass, Georgia, and Grand Prairie, Texas, adding more than 1,500 jobs.

The plants will build electrical equipment for data centers, which the group describes as systems for distributing, monitoring and controlling power. The group aims them at data centers built for AI.

The same release reports around €6 billion in data center orders over the first nine months of the fiscal year, with triple-digit growth in those orders. It also recalls €300 million announced in July 2026 for the group’s German plants.

ABB and power stability

ABB works on a different function: power stability at AI sites.

In November 2025, ABB announced it would supply Texas-based VoltaGrid with 27 synchronous condensers with flywheels and prefabricated electrical houses (eHouses), for AI sites under construction in the United States. Deliveries were due to start in December 2025, with the first units expected to be operational in April 2026.

In March 2026, the two companies extended the agreement to 35 more synchronous condensers, with the related eHouses and medium- and low-voltage distribution. Orders were expected in the second quarter of 2026. No financial details were disclosed.

The scope of these announcements

Siemens and ABB are long-standing electrical equipment suppliers. Their announcements concern production capacity and offerings built for AI sites.

Both examples are from the United States. On their own, they say nothing about the availability of the same equipment in Europe.

Two layers, two dynamics

On the compute layer, what is new is an additional supplier. Qualcomm is entering a segment held by Nvidia and AMD, with a product, a timeline and a first customer.

On the power layer, what is new is a change of scale at suppliers already in place. Siemens is adding manufacturing capacity; ABB is delivering stabilization equipment in volume to a single customer.

The two sets of announcements meet at rack power. Qualcomm announces 160 kW per liquid-cooled rack. The equipment Siemens and ABB sell is what distributes, controls and stabilizes power at that level inside the site.

That is why we track the two layers together. An inference rack can only go live once its power and cooling are in place, and those come from different suppliers on different timelines.

What these announcements change for GPU capacity planning

Availability dates to track on the compute side

For inference, the accelerator market now has one more announced option, with two milestones: 2026 for AI200, 2027 for AI250.

Both dates fall within the 18 to 36 month horizon over which GPU capacity is planned. They remain dates announced by the manufacturer.

Rack power gets a published reference

The 160 kW per rack figure announced for Qualcomm’s solutions gives a published order of magnitude for a rack-scale inference system in 2026.

The figure describes one product. It sets a power and liquid cooling requirement at rack level, which the hosting site has to be able to deliver.

The missing data

Three pieces of information are missing from the sources we reviewed.

The first is manufacturing lead time for electrical equipment. Neither Siemens nor ABB publishes it in these announcements.

The second is price, for accelerators and equipment alike. ABB states explicitly that financial details were not disclosed.

The third is independent measurement of the performance Qualcomm has announced.

Until that data exists, these announcements show where supply is heading without allowing anyone to put a number on the effect on project timelines or costs.

FAQ

Can Qualcomm replace Nvidia for inference?

The sources we reviewed do not support that claim. Qualcomm announces products built for inference, availability in 2026 and 2027, and a first customer for 200 MW. No independent performance measurement was available in these sources. The first deployments will settle the question.

Do Qualcomm racks require liquid cooling?

Yes, according to the manufacturer’s announcements: AI200 and AI250 are presented with direct liquid cooling, at 160 kW per rack.

Do Siemens’ investments cover Europe?

The more than $200 million announced on 7 August 2026 covers two US plants. The same release recalls €300 million announced in July 2026 for Siemens’ German plants, and the sources we reviewed do not say what share goes to data centers.

Planning GPU inference capacity in France over the next 18 to 36 months? Let’s talk about rack density and power for your deployment. [Contact VOLTEKKO]

Sources

This article might interest you as well Data Center Waste Heat: Three Conditions to Check Before You Believe It