跳到正文
The Next Platform·· 8 天前AI 評分44

Cerebras 籌備 CS-5 系統於 2027 年推出,並再獲一家超大規模雲端服務商合作

Cerebras Readies CS-5 Systems For 2027 Launch As It Bags Another Hyperscaler

AI 導讀

Cerebras 正籌備於 2027 年推出 CS-5 系統,並與 OpenAI、Amazon Web Services、Oracle Cloud 及 Microsoft Azure 等客戶簽訂合約。新一代 WSE-5 已提前完成晶圓製造並進入系統測試,CS-5 的速度將超過 CS-4 的兩倍。

正文

Upstart waferscale AI compute maker Cerebras Systems has been very upfront for all of 2026 that this was a year to build out its cloud and to ramp up its manufacturing capability so it can start meeting the demands of the contracts it has inked with OpenAI and now Amazon Web Services, Oracle Cloud, and Microsoft Azure. Thus far, most of its revenue backlog and all of its incremental revenue is coming from that OpenAI deal, which is finally ramping.

Next year is going to be a banner year for many AI system providers, and there is a non-zero chance that this could be the peak spending year for the GenAI boom. We shall see, and even AI systems do a terrible job at predicting the future. All anyone knows is what their customers tell them, and all those organizations know is what their customers tell them, and so on down and up the chain of being. Bullshit sometimes begets even bigger bullshit, but sometimes it isn’t bullshit and people really do buy a ton of systems. Which has happened for the past four years now despite the fears that the gravy train might run out of coal.

NEXTPLATFORM AD

Here is the thing, though: This GenAI thing is real. It is just a matter of time, driven by ever-decreasing cost of token chewing and spitting, before it becomes pervasive in the world’s applications, or around the world’s applications. Some go too far and think as the world’s applications, and that is because they do not understand the nature of risk avoidance by the millions of corporations around the world. There is mainframe code older than me at the center of our banking and insurance systems, and wave after wave of technological evolution since I entered the IT sector four decades ago has augmented that code but a very large portion of it is still there, like a granite mountain.

In the third quarter, Cerebras didn’t sell as much of its CS-3 and CS-4 hardware as it has done in prior quarters, and that stands to reason because the company was building out its cloud services business in support of the $10 billion inference deal that the company inked with AI model builder OpenAI back in January 2026.

Since then, that deal, for 750 megawatts of gear that OpenAI can rent for fast inference, has been expressed as being worth $20 billion and sometimes $30 billion between 2026 and 2032, inclusive. That larger revenue stream is apparently possible if OpenAI has Cerebras build an additional 1.25 gigawatts of capacity by the end of 2030 and the contract for services runs out to 2032. Provided OpenAI goes public and has the cash, there is no reason to believe that it will not stretch this deal to the max. This option by definition has to be cheaper than buying Nvidia GPUs for inference prefill and Nvidia Grok LPUs for decode. Particularly with Cerebras using AMD GPUs for prefill, as it is doing in its cloud.

In the third quarter, hardware sales at Cerebras were down 28.8 percent to $68.9 million, and because the Cerebras factory is still ramping to a new production level gross margins were pretty bad at only 11 percent of revenues instead of the long-term goal of 60 percent. Andrew Feldman, co-founder and chief executive officer at the company, told Wall Street analysts in a conference call that margins in Q3 2026 were also adversely affected by the aggressive pricing that an early customer – presumably cloud provider G42 – negotiated years ago for machines being installed now.

NEXTPLATFORM AD

The has been some jawboning out there on the intertubes that Cerebras has a yield problem, and Feldman brought the boot down on that idea. And it stands to reason. The WSE-3 used in the CS-3 systems and the WSE-3 Turbo used in the CS-4 systems are based on very mature 5 nanometer processes from Taiwan Semiconductor Manufacturing Co – nothing crazy like the 2 nanometer processes that AMD is spearheading with its “Venice” Epyc 9006 server CPUs, just launched this week. The WSE-3 Turbo is a stop-gap, double-clocked waferscale processor that Cerebras put into the market in August in the CS-4 systems and there is a WSE-5 (you can’t have a WSE-4 because there isn’t going to be one) in the works for next summer for the CS-5 machines. These CS-4 and CS-5 systems will compete with new GPUs from Nvidia and AMD and LPUs and other XPUs from Grok (essentially owned by Nvidia), SambaNova, Rebellions, and the homegrown gear from the hyperscalers, cloud builders, and AI model builders themselves.

“Our product road map continues to advance rapidly,” Feldman explained on the call. “Our CS-4 will soon be in full production, doubling both performance and throughput. Our next-generation processor, the Wafer Scale Engine 5, which will be deployed in our fifth-generation CS-5 system, is back from the fab ahead of schedule. The WSE-5 taped out on June 15. The chip is back, has been brought up, and is already running models and is undergoing in-chassis testing. Moving from tape-out to system test in less than five months is exceptional execution. The CS-5 will more than double the speed of CS-4. It will enter production in early summer of 2027. It is built on TSMC's 5 nanometer node, where competition for wafers is less intense and wafer costs are much lower.”

It is safe to say that that WSE-5 has been in development for a long time, and I strongly suspect that as technology has advanced, it has been tweaked to incorporate new things. I think there are all kinds of technological possibilities for this WSE-5 engine, including silicon photonics optical connections linking wafers and memory servers and SRAM stacking – two that I favor for the sake of Cerebras. At Hot Chips this past summer, Cerebras revealed that it would be using 3D stacked DRAM with its wafers.

I also think that given the need for fast inference, the wafers will be double-clocked from this point forward. Performance trumps power consumption and heat dissipation these days – dollars per millions of tokens is what matters more than dollars per millions of tokens per watt at this stage of GenAI commercialization. Electricity is cheaper than the systems, even if it is getting harder and harder to come by. Cerebras can always offer two different clock speeds at the same price and let customers decide.

NEXTPLATFORM AD

We shall see next summer. . . .

Given the lumpiness of hardware sales at all supercomputer makers – and make no mistake, the CS family is comprised of supercomputers – it is best to look at the company’s financials on an annual, not quarterly basis. Here is how 2022 through 2025 and projections for GAPP revenue (mine), not the non-GAAP revenue (which is larger) that Cerebras talks about in its financial presentations for Q3 and Q4:

And for those of you who like the raw numbers, here are the quarterly and annual results in the typical TPM monster table:

My best guess is that on a GAAP basis, Cerebras will close 2026 with around $885 million in revenues, up 73.4 percent compared to 2025’s sales. Operating income and net income were whacked by a spike in research and development spending and stock-based compensation. Feldman said on the Wall Street call that the company expected to triple its sales in 2027, and it is reasonable to assume that as manufacturing ramps and Cerebras increases its margins on hardware with partners Flex and Sanmina that gross profits and operating income will be much improved in 2027.

NEXTPLATFORM AD

Cerebras ended the third quarter with $7.36 billion in cash and equivalents in the bank and a revenue backlog of $26.1 billion, mostly from the OpenAI deal. The AWS, Oracle, and Microsoft deals are not yet in the backlog and are certainly not yet bringing in revenue. Cerebras has been pretty vague about what these deals are, but it is important to note that three of the biggest clouds in the world are going to be working CS iron into their inference workloads, presumably for decode operations. Whatever WSE-5 and the resulting CS-5 system is, this is what they almost certainly want, especially given that WSE-5 will have 2X the performance of the WSE-3 Turbo compute engine that is available now.

來源:The Next Platform · nextplatform.com