Silicon
-
From Fugaku
Fujitsu is selling its MONAKA inference CPU into the US server channel from 2027
Nikkei reports that an unnamed major US server maker has already agreed to buy Fujitsu’s MONAKA chip.
Nothing to do now; note it if you are writing a multi-year inference procurement that assumes one supplier.
-
The seller’s terms
Huawei raised its newest AI accelerator by up to half in two months, because it cannot buy the memory
Huawei now quotes its Ascend 950DT above ¥250,000, about $37,255, Reuters reported on 10 September. That is between 20 and 50 per cent more than it quoted customers two months ago. The constraint is memory, not the chip. Since Washington restricted exports of advanced high-bandwidth memory to China in December 2024, Chinese chipmakers have bought HBM, the stacked memory packaged beside an accelerator’s compute die, through grey-market channels at several times the price paid outside China.
If a deployment plan assumes Ascend or Cambricon parts at spring prices, re-quote it before committing. The input that moved is memory supply, not fab capacity.
-
In brief
Apple’s A20 Pro claims 50 per cent more memory bandwidth, which is the number that decides local decode
Apple states “50 percent more memory bandwidth than A19 Pro” and “double the AI processing power” from 32 Neural Engine cores. Local decode is memory-bound, so the bandwidth figure is the better proxy for tokens per second. The 32 cores arrive as a “Dual 16-core Neural Engine,” two blocks rather than one accelerator, which is a scheduling constraint. The release gives no RAM capacity, no absolute bandwidth, and no mention of on-device model sizes.
Size local-model plans against bandwidth rather than core count, and wait for an absolute figure before comparing this part to anything non-Apple.
-
Silicon
The first outside numbers on Google’s TPU put it a fifth cheaper than a B200 per token
SemiAnalysis ran the same open model through an unreleased TPU stack on Ironwood and through Nvidia’s B200 and B300. The TPU loses on raw throughput at most latency targets and wins on cost, and the software that made it possible is private until October.
If you are pricing 2027 inference capacity for an open-weight model, put Ironwood in the spreadsheet at roughly a fifth under a B200 per token at moderate interactivity, and mark it unavailable to you until the TorchTPU stack is public.
-
The number that moved
Nvidia now books $40B of revenue per gigawatt, against $18B on Hopper
Jensen Huang says Nvidia has stopped selling a chip and started selling a share of a datacentre’s electricity budget. The ratio he steers by more than doubled in one generation.
If you are budgeting 2027 inference, price it against a power envelope rather than a chip. Capacity is now rationed by megawatts agreed with utilities years ago, and no amount of demand moves that line.
-
Custom silicon
Broadcom’s AI chip revenue tripled, and none of it is a GPU
Broadcom’s AI semiconductor revenue reached $16.7 billion in the quarter to 2 August, against $5.2 billion a year earlier. It guided the next quarter to $21.7 billion. Total revenue was $29.6 billion, against $15.9 billion a year ago.
The alternative to waiting for Nvidia allocation is increasingly a part your cloud designed itself, which you reach through that cloud’s API and not through CUDA.