Culper Shorts Cerebras, Alleging 1.5-2.5 Year Chip Lifespans

Culper Research, a short seller, published a report on 28 October 2026 saying it is betting against Cerebras Systems because it believes the company’s wafer-sized AI chips “last just 1.5 to 2.5 years in the field”. Cerebras shares fell after the report; an Investing.com headline put the drop at 4%.

The report also points to a change in Cerebras’s own filings, which we checked. In plain words: how long a company assumes its machines last sets how fast it writes their cost off, and Cerebras raised that assumption between its December 2025 and April 2026 IPO filings.

Short explainer · AI chips brief

A vertical news Short narrated from the published headlines and opening lines of this piece and four related FourWeekMBA pieces on AI chip lifespans and depreciation.

Business Pill · Gross Margin

An evergreen Business Engineer explainer on gross margin. It explains the general idea only and does not discuss Cerebras or Culper.

What Cerebras’s Filings Say

In the draft registration statement Cerebras filed with the SEC on 22 December 2025, the useful life for “Computer equipment” is 1.5 to 3 years. In the Form S-1 filed on 17 April 2026, the category reads “Data center and computer equipment”, with a useful life of 3 to 5 years.

Both filings say: “Estimated useful lives are periodically assessed to determine if changes are appropriate.” The April S-1 adds: “Such revisions may result, for example, from changes to plans, demand, or strategy for the Company’s inference solutions.” Culper says the company “appears to have never explained the reason for this change.”

• Equipment on the books. Cerebras’s quarterly report for the three months to 30 June 2026 shows property and equipment of $1,077.8 million before depreciation, up from $488.6 million at the end of 2025.
• Depreciation. Depreciation expense was $24.4 million in the June quarter, against $5.6 million a year earlier, the same report shows.

What Was Posted on X

Milk Road AI posted that Culper Research disclosed a short position in Cerebras and published a report claiming its wafer-scale chips last only about 1.5 to 2.5 years in the field.

What Culper Alleges

Culper says it is short Cerebras, and its disclaimer tells readers to assume it holds a position in the securities it covers. Its central claim is that Cerebras never solved the heat problem of a chip the size of a whole silicon wafer. Instead, Culper alleges, Cerebras “programmed its wafers to run” dummy operations “such that the wafers never actually cool down.”

• Power at idle. Culper writes that Harvard and Meta researchers found in April 2026 that “the CS-3 draws 80% of peak power when idle,” against 20% for AMD and Nvidia GPUs.
• Its own model. Using a standard chip-ageing formula, Culper estimates 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.”
• Customers. The report quotes unnamed customers, including: “Our first system broke down, and then there was an issue with the replacement… Two replacements were made.”
• OpenAI deal. Culper estimates the contract’s value to Cerebras at “between roughly break-even and negative $13.8 billion at a 2.5-year wafer life.”
• Share supply. Culper says insiders have sold more than $400 million of stock since the IPO, that 19.4 million shares unlocked on 28 October and that more than 50 million more unlock in early November.

How Other AI Clouds Book Their Hardware

Other companies assume longer lives for AI hardware, according to the filings and reports in our 21 October piece: six years for CoreWeave’s technology equipment, 5.5 years for most Meta servers and five years for a subset of Amazon servers. Nvidia chief Jensen Huang has said his GPUs last “eight to maybe more” years, while Michael Burry’s estimate is two to three years, as CNBC reported.

Cerebras’s new 3-to-5-year range tops out at the low end of the CoreWeave, Meta and Amazon figures. Culper’s 1.5-to-2.5-year estimate for the wafers sits close to Cerebras’s earlier 1.5-to-3-year range for computer equipment.

Connecting the Coverage

The question of how long AI chips last runs through our coverage. Jensen Huang argued GPUs last eight years and could back AI loans, as we reported on 21 October. A 9,600-GPU A100 cluster listed on eBay for $59.7 million, about $6,220 per GPU including servers by our arithmetic, as we reported on 28 October. Cerebras’s compute deal with OpenAI, which Culper now values at break-even or worse, was detailed in September.

Burry flagged Anthropic notes that can pay rent in stock, as we reported on 28 October. For the basics of why chip lifetime drives AI profits, see our explainer. What’s new here: a short seller has put a specific lifespan on Cerebras’s chips and tied it to a change in the company’s own accounting.

The Structural Read

As we read it, the filings matter because Cerebras now sells cloud compute: Culper says that when a cloud system fails, Cerebras can record the replacement as capital spending, depreciated over the longer useful life.

Cerebras’s property and equipment more than doubled in the first half of 2026, to $1,077.8 million before depreciation, its quarterly report shows, so the useful-life assumption now applies to a much larger asset base.

Culper says it is short the stock, and its disclaimer says to assume it holds a position, so its estimates come from a party that profits if the shares fall.

Three Implications

FOR CEREBRAS INVESTORS The useful-life range in the filings sets the depreciation charge; Cerebras’s S-1 says lives are periodically assessed and may change with plans, demand or strategy for its inference solutions.

FOR AI CLOUD LENDERS Huang has said GPUs last “eight to maybe more” years and could back AI loans, as we reported; Culper’s estimate for Cerebras’s wafers is far shorter.

FOR OPENAI’S DEAL Culper estimates the OpenAI contract’s value to Cerebras at between roughly break-even and negative $13.8 billion at a 2.5-year wafer life; that is Culper’s model, not a filed figure.

The Business Engineer Lens

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

The framework states: “Inference accounts for 80 to 90 percent of the lifetime cost of a production AI system”.

As we read it, that makes the life of inference hardware a cost question as much as an accounting one, which is why a dispute over how long Cerebras’s wafers last bears on what its inference cloud earns. The framework was published in March 2026 and does not discuss Cerebras or Culper.

What Is Not Established

The 1.5-to-2.5-year lifespan is Culper’s estimate, built on its own model, customer interviews and former employees it does not name; we have not verified the customer accounts. Culper is short the stock and profits if it falls. We did not contact Cerebras for comment. The filings show the change in the useful-life range; they do not say how long Cerebras’s wafers last in the field.

The Bottom Line

Culper Research says it is short Cerebras and alleges its wafer-scale chips last 1.5 to 2.5 years. Cerebras’s filings show it raised the useful life of its computer equipment from 1.5-to-3 years in December 2025 to 3-to-5 years in April 2026, a range that tops out at the low end of the five to six years reported for CoreWeave’s technology equipment and for Meta and Amazon servers.