In this briefing
- Google plans to make a long-term US$17 billion investment in its first data centre in Wyoming, and Google plans a 702-megawatt fuel-cell development for the project. Crusoe confirmed only that Google had acquired the site and taken over the pre-development work. Neither company announced that the project was operational.
- Fireworks has scheduled the start of European deployments of additional NVIDIA Vera Rubin and Grace Blackwell capacity for 2027. Fireworks has also placed Ultrafast in public preview. The supplier's claim of more than 1,500 tokens per second is still not an independent test result.
- Qualcomm has completed hardware validation of its first-generation High Bandwidth Compute silicon, but Qualcomm Dragonfly AI250 is not expected to provide samples until the first half of 2027. vLLM v0.33.0, by contrast, is already available as an open-source Release.
A site transfer and a power plan still precede data-centre operation
On 6 November, Crusoe announced that Google had acquired its data-centre site in south-eastern Wyoming. Crusoe limited the completed work to site selection, government approvals and a hybrid power strategy developed with Tallgrass Energy, and Crusoe described the transaction as the completion of its pre-development work. Crusoe did not disclose the transaction price, and Crusoe did not describe the site transfer as a completed or operational data centre. On the same day, TechCrunch described the project as unfinished. TechCrunch presented concerns about cost and schedule as media reporting rather than as a fact confirmed by both parties to the transaction.
On 5 November, Google announced a long-term US$17 billion investment to build its first data centre in Wyoming. Google said that the company would pay the full cost of the power used by the project and would cover infrastructure needs directly driven by its own growth. Google also disclosed a planned 702-megawatt commercial-scale fuel-cell development that would use natural gas and carbon capture and storage, while the data centre is planned to use air cooling. Google did not disclose the final operational capacity, the annual spending schedule or a formal opening date. The US$17 billion investment and the 702-megawatt plan therefore cannot be treated as computing capacity that has already been delivered.
For enterprise adoption, control of a site, government approval, a power contract, construction progress and saleable computing capacity are separate acceptance objects. A project transfer can reduce or redistribute some development risks, but a project transfer does not automatically prove schedule, cost, power reliability or service availability. When an enterprise evaluates long-term computing supply, the enterprise needs to record signed site and power arrangements separately from services that have been built, connected and made available for purchase.
Regional capacity plans and current service capabilities occupy different time horizons
On 3 November, Fireworks announced that the company would expand European capacity with NVIDIA Vera Rubin, Grace Blackwell, DSX AI factory and Spectrum-X Ethernet. Fireworks explicitly scheduled the start of deployments for this additional capacity for 2027. The announcement therefore indicates the direction of future regional supply, but the announcement does not prove that the related capacity is already open to customers. Fireworks did not provide specific regions, saleable capacity, prices, supply batches or service-level commitments in this announcement.
Fireworks also announced that Ultrafast had entered public preview on the NVIDIA Blackwell platform. Fireworks also announced multi-day prompt caching, background mode, higher rate limits and cold starts of less than one minute. Fireworks claims that Ultrafast exceeds 1,500 tokens per second, but that figure comes from supplier material. The material does not provide the model, quantisation, context, batch, concurrency or tail-latency conditions. The regional expansion plan, the public preview capability and the supplier performance result therefore require separate verification.
For enterprise adoption, an assessment of a regional service needs to record data residency, available regions, capacity commitments, rate limits, cold starts, cache duration and the complete bill. A hardware expansion plan for 2027 can influence a medium-term architecture choice, but the plan cannot replace evidence of current capacity. A public preview can support a bounded evaluation, but the stability, price and service commitment of a public preview can still change, while supplier throughput figures also need to be measured again with the target model and request shape.
Validated silicon and a released runtime provide different degrees of availability
On 2 November, Qualcomm announced that its first-generation Qualcomm High Bandwidth Compute silicon had completed hardware validation in a San Diego laboratory. Qualcomm said that the tests covered the full memory read-write path and near-memory computation at production-representative dimensions, and Qualcomm said that a data-centre product had taped out. Qualcomm also stated that its comparison used measured results from first-generation silicon, normalised to target operating conditions and compared with published competing-product specifications. Final performance may still vary with the product implementation, workload and system configuration.
Qualcomm expects Qualcomm Dragonfly AI250, which integrates High Bandwidth Compute, to provide samples in the first half of 2027. This timetable shows that Qualcomm is currently presenting validated silicon and a product-integration path, rather than a complete accelerator that is already in mass production or on sale. Meanwhile, the GitHub Release page for the vLLM project already provides a v0.33.0 tag and release assets. The vLLM Release state proves that an open-source runtime version is available, but the Release page cannot prove that the version satisfies a particular enterprise's production requirements for its target model, graphics processing unit, quantisation, concurrency and failure recovery.
For enterprise adoption, an enterprise needs different evidence thresholds for silicon validation, engineering samples, production hardware, an installable runtime and a production service. Validated silicon can reduce the risk that the core technology does not exist, but validated silicon cannot replace information about complete-system power, memory capacity, network expansion, framework support, price, supply and after-sales support. A runtime Release can begin compatibility and regression testing, but an enterprise still needs to bind the runtime version to drivers, model formats, hardware firmware, benchmark conditions and production incident records.
What to watch next
- Can Google publish the phased construction progress, final operational capacity, connection date and saleable-service date for the Wyoming project?
- Can Fireworks publish the specific regions, capacity, price and service-level commitments for the European expansion, together with the complete test conditions for the claim of more than 1,500 tokens per second?
- Can Qualcomm publish the power, memory, interconnect, framework support, price and supply information for Qualcomm Dragonfly AI250 at the sample stage, and can vLLM continue to provide reproducible compatibility results for different hardware and model combinations?
Sources and verification
Golden Data has edited this briefing from the public materials listed above. The original sources govern facts and figures. The enterprise relevance sections are Golden Data editorial analysis and do not constitute an endorsement of any third-party product.
← Back to AI Daily Briefing