Cerebras Says Under 0.5% of Wafers Retired, Disputing Culper

Cerebras says that since the introduction of its CS-3 system in 2024, “less than half of one percent of all CS-3 wafers deployed have been decommissioned”, in a blog post by chief executive Andrew Feldman that disputes Culper Research’s short report. Culper had estimated that the wafers “last just 1.5 to 2.5 years in the field”.

Cerebras announced the CS-3 on 13 March 2024, 968 days or about 2.65 years before 6 November 2026, by our arithmetic. That is longer than the 1.4-to-2.3-year range in Culper’s own model, which Cerebras says is wrong.

Business Pill · Gross Margin

A short explainer on gross margin. It teaches the general idea only and does not discuss Cerebras or Culper.

What Cerebras Says

Feldman writes that “Nearly every financial conclusion in that report, from our margins to the value of our OpenAI agreement, depends on that one number.” He adds: “That number is wrong.”

On heat, the post says the wafer “is qualified by TSMC, its manufacturer, for 10 years of continuous operation at 105°C, twenty degrees above our maximum”. It says the fleet’s average junction temperature “is less than 65°C and doesn’t exceed 85°C”.

Using those inputs in the Arrhenius equation, Cerebras says a wafer held at 85°C “every hour of every day would last” 33 years before heat wore it out, and one held at 65°C “would last more than a century”. The post adds: “No one will run a CS-3 for 33 years. The point is that heat induced failure is not what will retire it.” The 33-year figure is 14.3 times the top of Culper’s 2.3-year model estimate, by our arithmetic.

On the connector, Cerebras says: “we don’t use the PariPoser and have never shipped a single system that uses it”. On the “dummy operations” the report described, it says they are “one of the tools” that keep power changes gradual, and compares them with a “power burner” mode it says Nvidia’s B300 design includes.

On accounting, the post says Cerebras first estimated useful life at 1 to 3 years when its compute equipment was mainly for research labs, then moved to 3 to 5 years once it ran production systems for its own cloud. It says “Most cloud providers currently report AI equipment with 5-6 year useful lives, placing us comparatively on the more conservative end”, and that the estimates are part of KPMG’s annual audits.

What Was Posted on X

Cerebras chief executive Andrew Feldman posted on 5 November that Culper Research’s report argued Cerebras wafers wear out after about two years, that the number is wrong, and linked the company’s blog post. The piece uses the blog post as its main source.

Alexander Bittan, posting on X, summarised Feldman’s rebuttal on 5 November, calling it ‘incredibly bullish’. That is his view; the piece takes Cerebras’s figures from the company’s own blog post rather than from his summary.

In an earlier post on 5 November, Bittan argued that AMD should consider buying Cerebras and that a 40% premium would require less than 4% dilution for AMD shareholders. That is his opinion and arithmetic, which the piece has not checked.

What Culper Alleged

Culper’s report of 28 October, as we reported it at the time, alleged that Cerebras “programmed its wafers to run” dummy operations “such that the wafers never actually cool down”. Using a chip-ageing formula, Culper estimated that “Cerebras wafers age 3x to 3.5x faster than GPUs” and that “If a GPU lasts 5 to 7 years, wafers can be expected to last just 1.4 to 2.3 years.”

Culper also put the value of Cerebras’s OpenAI contract at “between roughly break-even and negative $13.8 billion at a 2.5-year wafer life”. Feldman’s post says the report scaled the economic life of GPUs by a physical ratio, which it calls “a logical error”.

Culper’s X account had not posted a reply to the Cerebras post as of 13:40 UTC on 6 November, by our search of its recent posts. Its website showed a browser check and could not be read.

What the Stock Did

Cerebras shares closed at $81.71 on 5 November, Nasdaq data show, down 31.5% from $119.31 on 27 October, the day before Culper’s report, by our arithmetic. The 5 November close was 11.0% below the 4 November close of $91.85, on the day Feldman posted. It was also the day 12 AI Stocks Lose $1.45T at Close as Nvidia Falls 9.3%, as we reported, so the move cannot be tied to the post alone.

Connecting the Coverage

How long AI chips last is the question under several of this week’s stories: Culper Shorts Cerebras, Alleging 1.5-2.5 Year Chip Lifespans, Nvidia’s $108.5B Guarantees Meet 6-Year Lives at AI Clouds and Microsoft Ties 75% of Its $45.2B Capex to Short-Lived Chips. Cerebras now separates physical life, which it says runs decades, from economic life, which it says ends when newer chips do more per dollar and per watt.

The Structural Read

As we read the Cerebras post, it separates three reasons silicon is retired, mechanical failure, heat and economics, and argues that economics, not heat, will retire its wafers; Culper’s report, as we reported it, scaled GPU economic life by a physical ratio, which Cerebras calls a logical error.

As we read the field-data claim, the CS-3 has been on the market for about 2.65 years, by our arithmetic, so a 1.4-to-2.3-year life would imply a sizeable share of early units already retired; Cerebras says under half of one percent have been decommissioned, without giving the number deployed.

As we read the accounting section, Cerebras books its data center and computer equipment over 3 to 5 years, below the 5-to-6-year lives it says most cloud providers report, so its own books assume a shorter life than its heat model.

Three Implications

FOR CEREBRAS’S MARGINS Feldman says nearly every financial conclusion in the Culper report, from margins to the value of the OpenAI agreement, depends on the wafer-life number.

FOR CULPER Culper’s X account had not posted a reply to the Cerebras post as of 13:40 UTC on 6 November, by our search.

FOR THE GPU-LIFE DEBATE Cerebras says most cloud providers report AI equipment with 5-to-6-year useful lives, and books its own over 3 to 5 years.

The Business Engineer Lens

This story maps onto the Business Engineer framework The State of The Inference Economy.

The framework says: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”

As we read it, how many years a chip keeps earning that revenue is the number Culper and Cerebras dispute, and Feldman says the report built its margin and OpenAI-contract conclusions on it. The framework does not discuss Cerebras or Culper.

What Is Not Established

The 33-year and century figures are Cerebras’s model estimates for heat-induced failure only; the post says mechanical failure and economics also retire silicon. The under-0.5% figure is Cerebras’s own field data and does not say how many wafers have been deployed or replaced under warranty. Culper’s 1.4-to-2.3-year range is its own model. We did not contact either company.

The Bottom Line

Cerebras says under 0.5% of CS-3 wafers have been decommissioned since 2024 and puts heat-induced wafer life at 33 years at its 85°C maximum, against Culper’s estimate of 1.4 to 2.3 years. The CS-3 has been on the market for about 2.65 years, longer than the top of Culper’s range, by our arithmetic.