The easy question about this morning's Huawei chip story is how much the price went up and whether it undercuts NVIDIA. Both halves fail, for different reasons.
The second half fails because there is nothing to check it against. NVIDIA does not publish a price for the B200. Any sentence of the form "Huawei matched the B200" is a claim about a number that exists only in third-party estimates, and the estimates disagree with each other by more than the amount under discussion.
The first half fails because the increase has two different answers in circulation, and they are being run against each other as though one newsroom must be wrong. They came from two separate exclusives, with different sources, different windows, and different definitions of what was being measured.
So the question worth asking is narrower and duller: what is actually known about this price, who knows it, and what the money buys on either side of the comparison. The first two answers are one yuan figure, sourced to unnamed people at two newsrooms, for a card its maker says is not available until next quarter. The third is published by both companies, and it is not close.
Two exclusives, three hours apart
Reuters filed first on 10 September, reporting that Huawei lifted the indicated price of its Ascend 950DT accelerator to above 250,000 yuan ($37,255), a jump of 20% to 50% on quotes given to customers about two months earlier, "depending on contract terms", per two people. Bloomberg followed roughly three hours later with a suggested price up roughly 60% to 250,000 yuan ($37,300) over the past three months, per people familiar.
As each newsroom reported it on 10 September 2026 - no figure here is a Huawei disclosure
| Reuters | Bloomberg | |
|---|---|---|
| Filed, US Eastern | 01:10 | 03:41 |
| Price | Above 250,000 yuan | 250,000 yuan |
| In dollars | $37,255 | $37,300 |
| Increase | 20% to 50% | Roughly 60% |
| Window | Two months | Three months |
| Attribution | Two people | People familiar |
| Mentions the B200 | No | Yes, one person |
Reuters filed at 01:10 US Eastern, Bloomberg at 03:41 US Eastern. Reuters was read through its Investing.com and Yahoo syndications because the original is paywalled; both carry identical figures. Reuters describes an indicated price and a range that varies with contract terms; Bloomberg a single suggested price. The dollar figures convert the same 250,000 yuan at 6.7105 and 6.7024 respectively, both inside the 6.705-6.711 spot range of 9-10 September. Neither article prints a starting price.
The first thing the table settles is the dollar figures. $37,255 and $37,300 are the same 250,000 yuan, converted at 6.71 and 6.70 to the dollar — both inside the 6.705–6.711 range the yuan traded in on 9 and 10 September. The $45 between them is a rounding choice, not a second data point, and anyone quoting both is quoting one number twice.
The second is that only Bloomberg mentions the B200. Reuters' article does not contain the comparison at all — not in either of the two independent syndications we read it through. That matters, because the comparison is the part the market has fixed on.
The 50% and the 60% are not in conflict
Neither wire prints a starting price, so the bases have to be backed out of the percentages, and they are ours rather than anyone's disclosure. Bloomberg's 60% over three months implies a base near 156,000 yuan. Reuters' 20%-to-50% over two months implies a base between 167,000 and 208,000 yuan.
Those are not rival claims about one number. They are consistent with a single path — roughly 156,000 yuan three months ago, somewhere in the high 160,000s to low 200,000s two months ago, 250,000 now — requiring a rise of between 7% and 33% in the first of those three months. Two windows, one price path, no contradiction.
There is a caveat that limits even that reconciliation, and it is the more important point. Reuters is describing a spread across contracts: 20% to 50% "depending on contract terms". Bloomberg is describing a single suggested price. Those are different kinds of quantity, and compounding one into the other assumes they measure the same thing. They may not. Either way, the argument the tape is having — 50% or 60%, which wire got it right — is about a disagreement that does not exist.
The comparison is to a price that has never been published
Bloomberg's parity line is precise about its own weakness: the price "was set in line with Nvidia's B200, according to one of the people". One person, unnamed, characterising an internal pricing decision at a company that has commented to neither wire.
Set that against the thing it claims parity with. NVIDIA publishes no list price for the B200; every per-GPU figure in circulation is a reseller quote or a system price divided by eight. Thunder Compute's pricing survey, last reviewed on 4 September, puts three of them side by side.
US$ per GPU, against the 250,000 yuan indicated for the Ascend 950DT
| B200 basis | Per GPU | $37,300 sits |
|---|---|---|
| List, 8-GPU volumes | $30,000-$40,000 | Inside, near top |
| Street, single SXM | $45,000-$55,000 | 17% below |
| 8-GPU HGX system | $56,250-$68,750 | 34% below |
Bands are Thunder Compute's survey of reseller quotes and system-implied prices, last reviewed 4 September 2026, which states that NVIDIA does not publish list prices. The system row divides its quoted $450,000-$550,000 eight-GPU HGX B200 price by eight, and so carries CPUs, networking, chassis and margin that the bare-accelerator rows do not - it is the least like-for-like of the three. The comparison column is ours, measured against $37,300 and stated relative to the nearest edge of each band.
On the volume-list basis the parity claim is defensible — 250,000 yuan lands inside the band, near its top. On the street basis the Huawei card is about 17% cheaper. On a fully-built eight-GPU system basis it is about 34% cheaper. The claim is true, roughly true, or plainly false depending on which unpublished number the unnamed person had in mind, and nothing in either article says which.
What would settle it is a document — a Huawei price list, a distributor quote sheet, an IR statement carrying a 950DT figure. None exists publicly. Both wires rest on people, and so does everything downstream of them.
The same money buys about a quarter of the compute
Price parity, even taken exactly as the unnamed person put it, is a statement about one number. What the number buys is the comparison underneath it, and here both companies have published enough to settle it without anyone's sources.
Huawei's own figures for the 950DT, given by Eric Xu on stage: 1 PFLOPS of FP8, 2 PFLOPS of MXFP4, 144 GB of HiZQ 2.0 memory at 4 TB/s. NVIDIA's published DGX B200 specification, divided by the eight GPUs it describes and halved to the dense basis its own footnote defines: 4.5 PFLOPS of FP8, 9 PFLOPS of FP4, 180 GB of HBM3e at 8 TB/s.
Per accelerator, dense basis, from each maker's own published figures
| Published spec | Ascend 950DT | B200 |
|---|---|---|
| FP8 compute | 1 PFLOPS | 4.5 PFLOPS |
| FP4 compute | 2 PFLOPS | 9 PFLOPS |
| Memory | 144 GB | 180 GB |
| Memory bandwidth | 4 TB/s | 8 TB/s |
| Price per GPU | $37,300 | $30,000-$40,000 |
| Per FP8 PFLOPS | $37,300 | $6,667-$8,889 |
Ascend 950DT figures are Huawei's own, from Eric Xu's Huawei Connect keynote of 18 September 2025: 1 PFLOPS FP8, 2 PFLOPS MXFP4, 144 GB of HiZQ 2.0 at 4 TB/s. B200 figures are NVIDIA's published DGX B200 specification - 1,440 GB and 64 TB/s across eight GPUs, FP8 Tensor Core 72 PFLOPS and FP4 144 PFLOPS - divided by eight and halved, because NVIDIA footnotes those two as sparse and defines dense as half. Huawei's keynote does not say whether its PFLOPS are sparse or dense; read here as dense, which is the reading most favourable to Huawei. The B200 price band is the volume-list row of the previous table, the only basis on which the parity claim holds. Both per-PFLOPS lines are ours, dividing price by FP8 compute.
Run the price through that. On the volume-list band — the one basis where the parity claim holds — the Huawei card costs $37,300 per FP8 PFLOPS against $6,667 to $8,889 for the B200: 4.2 to 5.6 times more for the same unit of compute. Per terabyte per second of memory bandwidth the gap is narrower, 1.9 to 2.5 times. Even against the street band, where the 250,000-yuan figure looked 17% cheaper than a B200, it is 3.0 to 3.7 times more expensive per PFLOPS.
One caveat on the basis, and it runs against Huawei. NVIDIA marks which of its figures carry sparsity and which do not; Huawei's keynote marks neither. Read as dense, the gap is what the table shows. Read as sparse, the dense figure halves and the gap doubles.
None of this makes the 950DT a poor purchase for a buyer who cannot legally take delivery of the alternative — that is the whole point of a domestic substitute, and scarcity is what a price under export controls measures. It does mean "set in line with the B200" describes the invoice and not the machine.
What the price is attached to
Two facts limit how much any of this can mean.
The card is not available yet. Huawei's own roadmap, given by Eric Xu at Huawei Connect in September 2025, puts the Ascend 950DT in the fourth quarter of 2026 — a window that opens three weeks after these reports. Every figure here is indicated or suggested, for a part nobody has yet bought at any price.
And the reason both wires give for the rise is a cost, not a win. The driver they name is the HBM shortage, sharpened by the US controls on advanced memory exports to China in force since December 2024, which have pushed Chinese buyers into grey-market channels where Reuters reports the memory "typically costs several times what buyers outside China pay". Huawei is not the only one: Reuters has Cambricon's planned 690 up 20% to 30% on two-month-old indications, with MetaX and Iluvatar CoreX moving similarly, and Huawei's own older parts moving on a third time window again — the 950PR to more than 80,000 yuan and the 910C to more than 110,000, from roughly 60,000 and 90,000 at the start of the year.
That is a domestic substitute getting more expensive because its inputs are scarce. It reads across to the memory suppliers rather than to NVIDIA, and to the extent it reads across to NVIDIA at all it cuts the opposite way from how the story has travelled: the alternative Beijing is pushing customers toward just got dearer. We have bounded an unpublished NVIDIA chip price once before, and the discipline is the same — the number you cannot check is the one carrying the argument.
Two things would change this. A published Huawei price sheet would replace the anonymous sourcing with a document. An NVIDIA list price for the B200 would make the parity claim testable. Neither is likely, which is the honest state of it: an indicated price on an unshipped card, benchmarked by one person against a number nobody outside NVIDIA has seen.
The Ascend 950DT, 950PR, 910C and Cambricon 690 price figures, the source counts, the contract-terms qualifier and the grey-market HBM characterisation are Reuters' reporting of 10 September, read through two syndications because the original is paywalled; the roughly-60% figure, the three-month window and the B200 parity line are Bloomberg's separate exclusive the same morning. Both rest on unnamed people, and Huawei, Cambricon, MetaX and Iluvatar CoreX all declined to comment or did not respond. The 950DT availability quarter and specifications are Huawei's own, from Eric Xu's Huawei Connect keynote of 18 September 2025. B200 price bands are Thunder Compute's survey of reseller and system-implied quotes, last reviewed 4 September 2026; NVIDIA publishes no list price. B200 compute, memory and bandwidth are NVIDIA's own published DGX B200 specification, divided by the eight GPUs it describes and reduced to the dense basis NVIDIA's footnote defines. Dollar-yuan spot is the 9 and 10 September market rate. The implied base prices, the per-GPU system division, the currency reconciliation, the per-PFLOPS and per-terabyte-per-second divisions and every percentage gap between the Huawei figure and the B200 bands are ours.
