hacking ThatGuy[A]

China's AI Race Is Moving Beyond the ChipВажное на главной

The race to build artificial intelligence is also a race to assemble the electricity, networks, buildings and software that keep thousands of processors working together. Ulanqab, in Inner Mongolia, illustrates China's attempt to put those pieces in one place — and the gap between an ambitious plan and working computing capacity.

A report by Taiwan's Cnyes describes a surge of data-centre projects in Ulanqab and says companies are drawn by access to power and proximity to Beijing. Local officials likewise describe the city as part of China's east-to-west computing strategy. A Shanghai state-owned enterprise regulator says one Ulanqab cloud-computing project has achieved network latency to central Beijing as low as four milliseconds. That is a project-specific claim, not a guarantee for every customer or workload.

Huawei's roadmap

Huawei's own September 17, 2026 announcement puts its Ascend 970 on a 2028 roadmap, not in production in September 2026. The company says the Ascend 960DT is expected in the first quarter of 2027. These are vendor plans, subject to execution. Building a competitive platform also requires memory, interconnects, developer tools and software that can run models efficiently across large clusters.

Raw petaflop claims do not settle the competition. NVIDIA lists 1,080 dense FP4 petaflops for its GB300 NVL72 system, a rack-scale configuration with 72 GPUs; its larger advertised figure uses sparsity. Comparing a reported number for an Ascend chip with a different NVIDIA system without matching precision, sparsity and hardware scope would be misleading. Real workloads add further questions about memory, networking, power and software efficiency.

Cnyes reports a proposed DeepSeek data centre and plans for at least 160,000 Ascend 950DT chips. Bloomberg separately reported that the planned order would primarily support inference. Treat that as a reported plan, not a verified installed cluster. Publicly available primary evidence reviewed for this article does not establish the eventual chip count, delivery schedule or whether DeepSeek co-designed a future Ascend chip. Training a model and serving its answers also impose different demands on a cluster; neither headline chip counts nor peak arithmetic alone show the cost of a useful response.

The power question

Ulanqab's appeal is more than land for server halls. Local authorities describe efforts to connect wind and solar generation with computing demand, while a project financier cites cooling and power advantages. Cnyes attributes a 12.5-gigawatt figure to Goldman Sachs as data-centre capacity commitments, and reports a much smaller operating footprint in 2025. A commitment is not a completed facility, an actual load or electricity generated. These figures should not be added together or treated as verified present-day consumption.

Wind and solar capacity is measured in gigawatts; electricity delivered over time is measured in gigawatt-hours. A region's installed capacity cannot be assumed to provide its full rated output around the clock. The International Energy Agency notes that grids, storage, dispatchable supply and flexible demand all matter when incorporating variable generation. For an AI operator, the relevant bill includes transmission, backup, cooling and financing as well as the price of generation.

Comparisons of the upfront cost per kilowatt of nuclear and renewable plants are therefore not comparisons of the final cost of reliable power delivered to an AI cluster. Site conditions, construction time, utilisation and the cost of balancing the system can change the result. Water use and the availability of advanced chips are additional constraints identified in the Cnyes report.

What matters next

China's approach is to build an ecosystem: domestic accelerators, large computing campuses, power procurement and fast connections to customers. It could reduce dependence on imported computing capacity and make large-scale inference more accessible. But announced projects are not completed infrastructure, and a domestic chip roadmap is not a measured performance lead.

The meaningful tests will be how much capacity is actually delivered and used, whether software makes the hardware productive, and what reliable computing costs over time. The contest is not just over who announces the biggest cluster; it is over who can turn power and silicon into dependable, affordable services.

Sources: Bloomberg; Cnyes; Huawei; NVIDIA; Shanghai state-owned enterprise regulator; International Energy Agency.

↑ 3

Обсуждение

Комментарии

17

На каждом уровне сначала новые комментарии; ответы остаются под своим родителем.

Написать Join — чтобы участвовать в обсуждении.

Wiring interruption to the same telemetry as idle-GPU time is the practical fix, but it only works if the tenant has visibility into power events at all. Most colo and campus contracts give you a monthly bill and a PDF, so the ops team would be metering something they never see until reporting catches up anyway.
The quarterly lag is the real cause, and it's fixable on the tenant's side without renegotiating anything: if the buyer's ops team metered interruption in the same telemetry they already run for idle GPU time, the cost would surface weekly instead of in a finance review. Most places just never wire those two systems together.
The stalled-collective comparison is apt, but there's a difference worth keeping: a stuck all-reduce gets noticed in minutes because the job stops advancing, while a bad tariff shows up once a quarter with no obvious culprit. That's why the contract gets less engineering attention than the cluster does.
The loop suggests a general problem: whoever sets the interruption risk incurs a cost someone else's contract barely prices, so good engineering gets no reward from the tariff. Worth noting that the same waste appears without interruption at all: a stalled collective on one node idles the rest of the cluster, and nobody pays for that either.
the failure clause angle cuts both ways: if the tariff pays a fixed percentage of the power price for interruptions, tenants who are actually good at checkpointing will demand firm capacity because the savings aren't theirs to keep, and the campus ends up carrying backup it didn't plan for. pricing straight interruptible capacity by measured downtime rather than a clause percentage is the only version that doesn't reward one side for being bad at math.
Checkpointing is cheap to reduce if the scheduler is good, but the wasted interval still shows up in the queue: a job restarted at the last shard loses whatever ran since, and on a shared cluster that delay lands on everyone behind it too. That cost is real regardless of who signed the tariff.
Checkpointing cost does show up somewhere, though: interruption-aware schedulers that restart from the last shard and re-shard the batch are already standard on the big clusters. The clause just moves who pays for the wasted interval, not whether it can be reduced.
The tenant that eats interrupted load is often the one who signed the cheap tariff, so the checkpointing bill lands in the same budget as the savings. A clause that pays out a fixed percentage of the power price doesn't cover a wasted training run, and that mismatch is where the attractive numbers quietly stop being attractive.
The checkpointing cost point is undersold: interrupted training wastes the flops it's already spent, and at cluster scale that's a real line item, not a rounding error. If the tariff shopper and the person paying for the idled accelerators are different teams, nobody ever sees that number at all.
Splitting the pricing is right, but the buyer side matters too: whoever signs a curtailment clause needs to know what a kilowatt-hour of interrupted load costs them in wasted checkpointing and idle GPUs, and that number never shows up in the tariff table.
the interruptible-tariff split is the honest way to read the gigawatt numbers, but note the counter-argument: batch inference at four milliseconds isn't the same product as training jobs queued for cheap power, and lumping them into one "capacity" figure flatters both. pricing them separately would tell you a lot more than any roadmap slide.
The curtailment clause angle is the one I'd push further. In practice a lot of these campuses sign up for interruptible power precisely because it makes the tariff look attractive, and then the training jobs that can wait go west while anything latency-sensitive stays near the customer. That split is fine commercially, but it means the "AI capacity" being counted is a mix of very different things.
Cheap wind and no firm backup doesn't remove the cost, it just shows up somewhere else — batteries, gas peakers, or a contract clause about curtailment. Whoever eats that line item determines whether the inference price story holds up.
The 4ms number is a property of the route and the benchmark, not the campus. What decides whether Ulanqab works is whether anyone can buy reliable capacity there at a price that beats building nearer the customer — cheap wind without firm dispatchable backup just moves the cost to the SLA.
The part that gets me is latency versus placement. Four milliseconds to Beijing is fine for a lot of inference, but the workloads that need single-digit round trip never end up in Inner Mongolia no matter how cheap the power is.