Chips

Cerebras moves prompt processing to AMD GPUs, claiming 5x tokens per megawatt in production

With Q4 cloud revenue guided flat because Cerebras has no spare data center capacity, more tokens per megawatt is the only cloud lever it controls until new power lands from mid-2027. Whether its self-measured 5x holds decides how CBRS grows in between.

Richard TangRichard Tangthree-minute read
Share
Andrew Feldman, chief executive and co-founder of Cerebras Systems, in a dark suit jacket and purple shirt against a dark background
Andrew Feldman, Cerebras Systems chief executive and co-founder, on the November earnings call: capacity was too thin for a broad rollout of OpenAI's Ultrafast mode.

Cerebras said on November 17 that disaggregated inference is running in production in its Inference Cloud: AMD+4.57% — AMD, up 4.57 percent today Instinct GPUs process the prompt, its CS-3 wafer-scale systems write the response, and the company measured five times the throughput per megawatt of a wafer-only setup, a production number rather than the early-test claim of early October 1.

The launch matters this month because Cerebras−9.82% — Cerebras, down 9.82 percent today has run out of power to sell. Last week, CFO Robert Komin guided fourth-quarter cloud revenue roughly flat against the third quarter's $145 million: the ramp is "limited entirely by our data center capacity" 2. Until new megawatts arrive, the only thing Cerebras can sell more of is tokens per megawatt.

Each chip now does the half of inference it is built for

Every request has two stages. Prefill reads the prompt, a parallel compute job that suits GPUs. Decode emits tokens one at a time, rereading the model from memory at each step, and Cerebras holds model weights in on-chip SRAM for exactly that bandwidth job 1. Handing prefill to AMD frees wafer capacity for token generation 1. AMD announced the joint design in July and said Cerebras would deploy its Helios racks in its own data centers 3, so the company built on rejecting GPUs now runs its cloud on both.

The design targets the constraint that is capping the P&L

New capacity starts arriving in December, and the bulk of about 700 gross megawatts under contract lands only from mid-2027 24. Most of the $26.1 billion backlog sits with one customer, OpenAI, under a take-or-pay contract for 750 megawatts 2. Cerebras is behind on the first tranches and owes delay credits, and CEO Andrew Feldman said capacity was too thin for a broad rollout of OpenAI's$1.18T — OpenAI, private, latest valuation $1.18T Ultrafast mode 2. On the call he framed the new design in the same terms, as a way to produce "even more high-speed tokens when high-speed tokens are constrained" 2.

Cerebras now spends more than twice its revenue building capacity

  • Capital spending
  • Revenue
$0M$200M$400M$600MQ1 '25Q2 '25Q4 '25Q1 '26Q2 '26Q3 '26$497.19M
Data
Capital spendingRevenue
Q1 '25$98.24M$99.51M
Q2 '25$86.85M$103.32M
Q4 '25——
Q1 '26$131.97M$193.41M
Q2 '26$416.9M$180.11M
Q3 '26$497.19M$210.81M
Cerebras Systems revenue and capital spending by quarter, in USD millions, from its SEC filings. The filings retrieved report no figures for Q4 2025, so that quarter is left blank. Source: [8].8

The gain is self-measured, and the durability question stays open

The 5x figure is Cerebras' own, measured on November 14 on a single shared endpoint running the open GPT-OSS-120B model against its own baseline 1. The open question is whether it survives partners' accelerators and workloads as the design rolls toward hundreds of megawatts by end-2027, with AWS+3.47% — AWS, up 3.47 percent today, Fireworks AI~$1.1B — Fireworks AI, private, latest valuation ~$1.1B, Gimlet Labs and General Compute in early access 1.

Culper Research's October 28 short report claimed the wafers wear out in about two years and noted that Cerebras lengthened the systems' assumed useful life between draft IPO filings 5. Offloading prefill asks less compute of each wafer; whether that touches the durability claim, management has not said, and practitioners were still arguing it in a 92-point Hacker News thread on the launch 6.

The number to watch is cloud revenue per megawatt

Greenoaks, an $18 billion fund, sold its Cerebras stake in the third quarter 7. At about 93 times forward earnings, CBRS needs growth that cannot wait for concrete 7. The number to watch through the first half of 2027 is cloud revenue per megawatt on the fleet Cerebras already runs: if even part of the gain is real, the flat fourth quarter was a power problem, not a demand problem, and making scarce power produce more tokens is exactly the work the market is paying this company to do 7.

Get Chips by email, free

How AI demand is reshaping chipmakers, memory and the fabs behind them. In our daily email, every weekday.

Loading AMD

Reader comments

0 comments

    Chips

    Suzhou TFC posts 55% Q3 profit growth and buys Thai factory land as a US export rule names Thailand

    The Shenzhen-listed supplier's co-packaged-optics parts are already in batch delivery, while TE Connectivity's rival New York plant opens no earlier than spring 2027. A US export rule effective Thursday names Thailand, where TFC just bought factory land.

    Chips

    Cirrus Logic, which gets 93% of sales from Apple, guides December below last year's $580.6 million

    Cirrus Logic, which supplies the audio, haptic and power chips in Apple's iPhones, held September revenue flat at $559.5 million while its finance chief called the quarter's gross margin a high point. Management says the softer December guide is timing and KeyBanc says Apple cut wafer starts, and the December report decides which is right for a stock that takes 93 cents of every sales dollar from Apple.

    Chips

    Nvidia board rules out Huang return, starts permanent CEO search

    A unanimous board ruling and a hired search firm settle AI's biggest leadership question, on a record quarter reported six days into interim CEO Colette Kress's tenure. For holders and shorts, the risk has moved from who runs Nvidia to what Washington and Beijing let it sell.

    Chips

    B.Riley cuts AXT price target to $43 from $56 on China permit risk

    Washington's curbs on Chinese optical transceivers lifted AXT's biggest customers and pointed straight at AXT's single point of failure: all of its indium phosphide is grown in China and shipped on a Beijing permit. The stock now trades below both the $43 and $90 targets its covering analysts published this month.

    Chips

    SiTime launches FemtoClock 4 and VersaClock 8 into a $2 billion AI-rack market

    Both clock families ship now, and SiTime's own five-year sizing lands the same month its data-center segment missed consensus by $19 million and Goldman Sachs cut that segment's estimates. The December quarter, reported in early February, decides whether the launch or the line sets the multiple.

    Chips

    Nvidia adds Samsung Electro-Mechanics capacitors to its supply deal as prepayments top 1 trillion won

    Nvidia amended its supply deal with Samsung Electro-Mechanics to add MLCCs, but the dollar figures in its announcement went to Amkor, SPIL and Kinsus. Customer prepayments of $777 million are underwriting the Korean company's buildout instead, and most of those contracts expire in December 2027.

    Subscribe to TheInference for free