Why Cerebras Stock Is Falling Below Its $185 IPO Price in…

Cerebras Systems stock is falling because a customer-hardware report and an insider sale hit the same week. They are not the same fact. Nasdaq shows Class A shares, ticker CBRS, closed at $169.52 on October 1, 2026, down $8.13, or 4.58%. That close is 8.4% under the $185.00 IPO price in the May 14, 2026 prospectus, and 5.4% above the $160.81 52-week low. The break under $185 began on September 30, after Investing.com reported that SemiAnalysis had posted on X that OpenAI’s GPT-6.1 Sol Ultrafast tier was running on Nvidia GPUs at a low batch size, not on Cerebras chips. Chief Operating Officer Dhiraj Mallick also sold stock. His Form 144 notice, dated September 29, lists 396,000 shares at an aggregate market value of $77,909,040. The Form 4 filed September 30 says those shares sold on September 29. FinanceFeeds checked the filing on 2 October 2026. Premarket on October 2, Nasdaq showed $172.55 at 5:35 a.m. ET, still under $185.

A missing logo and a personal sale answer different questions. Mallick sold under a Rule 10b5-1 plan adopted on June 30, 2026, so the Form 4 does not reveal what he thinks of this week’s report, and the cash went to him, not to Cerebras. The hardware report is the item that can change revenue, and only for the tier it names, and only if it is true. If GPT-6.1 Sol Ultrafast is on Nvidia, Nvidia is the reported winner of that lane. Cerebras’s own page still says the company powers an earlier tier, GPT-5.6 Sol Ultrafast. The December 2025 contract for 750 megawatts of capacity is a separate document.

What happened to the Ultrafast tier

The move is two sessions. CBRS closed at $194.95 on September 29, at $177.65 on September 30, down 8.9% on 13,896,790 shares, and at $169.52 on October 1, down another 4.58% on 12,203,330 shares. That is a 13.0% drop. On October 1 the range was $168.40 to $184.67, so the session never traded back through $185.

The sourced headline is narrower than “Cerebras lost OpenAI.” OfficeChai, on September 30, reported that SemiAnalysis had written on X that GPT-6.1 Sol Ultrafast was not running on Cerebras and was running at a low batch size on Nvidia GPUs. Investing.com’s account, carried the same day by Yahoo Finance, says the same thing and names that tier. OfficeChai added that OpenAI had not said which hardware serves it. FinanceFeeds could not retrieve the original X post on October 2. It is also not a September newsletter. The last Cerebras feature in the SemiAnalysis archive we could find is an August 19 note on the CS-4, not a September 30 hardware call.

Batch size is the product point, if the report is right. A high batch keeps a GPU busy and cuts the cost per token. A low batch makes each user wait less and uses the chip less efficiently, which is why a speed tier can cost more. Cerebras has argued that a wafer, with the weights kept on the chip, is built for that job. A claim that Nvidia is doing it instead is a claim about one lane, not a cancellation of the 750-megawatt contract.

The earlier Ultrafast product is still on the company site, and it names Cerebras. The Cerebras blog says the company and OpenAI shared an early look at Ultrafast mode for GPT-5.6 Sol, launching first in the OpenAI API and powered by Cerebras, at up to 750 output tokens per second, in a limited preview. OfficeChai’s account of the newer tier put the speed nearer 300 tokens per second. Those are different products. Mixing them is how a hardware rumor turns into a story that the whole relationship moved.

The page says GPT-5.6 Sol Ultrafast is powered by the Wafer-Scale Engine, and that each wafer-sized chip carries 44 GB of SRAM so weights stay on the chip. In a May 13, 2026 newsletter, SemiAnalysis described the WSE-3 as fabricated on TSMC’s N5 process, and argued that fast-token tiers were why a lab would pay for wafer capacity. That is the opposite of the September post the stock is reacting to. Both can be in the record. Only the September one is the catalyst the recaps name.

What Cerebras, OpenAI, and Nvidia have said

The only on-the-record quotes are about the tier Cerebras says it already serves. They are not a reply to the September 30 report.

Rohan Varma, Product at OpenAI, is quoted on the Cerebras page: “With GPT-5.6 Sol Ultrafast, Cerebras enables AI that keeps up with how you think, code, and collaborate. We’re excited to see how workflows and applications are transformed by Ultrafast inference.”

Jeffrey Wang, an OpenAI researcher, is quoted on the same page: “Whereas formerly I might have to wait a couple minutes for a task to finish, it now finishes for me before I even have the opportunity to context-switch. It makes me way more productive.”

Both lines are about GPT-5.6 Sol Ultrafast. The page does not mention GPT-6.1 or a September hardware change. FinanceFeeds checked that page on October 2, 2026, and the GPT-5.6 claim was still up. We did not find a newer Cerebras sentence taking the GPT-6.1 tier or giving it up.

The September 30 recaps describe silence, not a confirmation. OfficeChai wrote that OpenAI had not named the GPT-6.1 hardware and that Cerebras had not commented. Investing.com said Cerebras had not confirmed or denied a future role. No later comment from Cerebras, OpenAI, SemiAnalysis, or Mallick settles it. Nvidia is the winner in the report. It is not yet the winner in a filing or on the company blog.

Nvidia did not have to confirm anything for its own close to look firm. Nasdaq shows Nvidia at $230.86 on October 1, up $2.48, or 1.09%, inside a 52-week range of $164.27 to $236.54. That is not evidence a speed tier changed hands. FinanceFeeds has laid out longer Nvidia scenarios separately. The open question is whether a low-batch GPU tier, if OpenAI is really running one, replaces wafer capacity or sits beside it. Off-chip memory is the bottleneck Cerebras names when it contrasts the wafer with a GPU. Micron’s latest quarter is a memory-supplier tape, not a line in the Mallick filing or in the Ultrafast recaps.

Where the stock sits against the $185 offer

Cerebras (CBRS) daily closing prices from May 14, 2026 through October 1, 2026. The dashed line is the $185.00 IPO price in the May 14, 2026 prospectus. Source: Nasdaq daily closes. Chart prepared October 2, 2026.

October 1 was not the first close under the offer. Of 97 Nasdaq sessions from the May 14 listing through October 1, 20 closed below $185. The stock closed at $168.52 on June 25. The $160.81 52-week low was the intraday low on June 26, and the shares still closed that day at $181.59. They closed at $169.39 on July 29. The latest official close, $169.52, is 5.4% above that low. Treating $161 as today’s price is wrong. What is new is the path. The stock closed at $206.75 on September 25 and at $194.95 on September 29, then gave that rebound back. Average volume over the 97 sessions was about 7.4 million shares. September 30 was 13.9 million and October 1 was 12.2 million.

ReferenceLevelVersus the October 1 close of $169.52

October 1, 2026 close$169.52Down 4.58% on the day
IPO price, May 14, 2026 prospectus$185.00Close is 8.4% below the offer
First-day close, May 14, 2026$311.07Down 45.5% from that close
52-week high, May 14, 2026$386.34Down 56.1% from the high
52-week low, June 26, 2026 intraday$160.81Close is 5.4% above the low
September 29, 2026 close$194.95Down 13.0% to the October 1 close
Nvidia close, October 1, 2026$230.86Nvidia rose 1.09% that day

On May 14, CBRS opened at $350, traded between $300 and $386.34, and closed at $311.07 on volume of 33,497,870 shares. The offer price in the Form 424B4 was $185.00. A buyer at the first-day close is down 45.5%. A buyer at the offer is down 8.4% on the October 1 close. Most of that gap was made between May and July, before GPT-6.1 was the story.

The 112,247,109 shares outstanding on the Form 144, times the October 1 close, imply a market value near $19.0 billion. That is the count on the notice, not a float figure. The June quarter, filed August 12, is why a customer headline can still move a growing stock. Revenue was $180.1 million, up 74% from $103.3 million. Six-month revenue was $373.5 million, up 84%. Quarterly gross margin was 14.2%, against 31.1% a year earlier. The GAAP net loss was $450.5 million, against net income of $309.5 million a year earlier. The June 30 10-Q is the source. Three customers were each at least 10% of the quarter, at 34%, 32%, and 10%. OpenAI arrangement revenue, disclosed separately, was $56.8 million.

The COO sale is a Form 4, and it is not revenue

The “$78 million” shorthand is the Form 144, and a Form 144 is a notice, not a sale. Mallick’s notice, signed September 29, 2026, proposes 396,000 shares through Morgan Stanley Smith Barney at an aggregate market value of $77,909,040. It splits them into 276,000 shares from options exercised that day and 120,000 performance units acquired on May 13, 2026, and it lists an August 18 sale of 38,889 shares for $8,642,077.14. The plan date on the notice is June 30, 2026.

The completed sale is the Form 4, signed September 30 for September 29 trades. His title on the form is Chief Operating Officer, the role the prospectus dates to September 2023. A same-day conversion took the reported Class A line to 396,000 shares, and 19 sale lots then sold all 396,000. Weighted-average prices run from $212.93 to $194.91. Lot size times those prices is $80,798,603.74. The notice had assumed a lower price. That Class A line ends at zero. Class B still converts into Class A, so the form does not say he owns nothing else.

One footnote separates the sale from the Ultrafast headline. The Form 4 says the shares were sold under a Rule 10b5-1 plan adopted on June 30, 2026. That date is before the September reports and before the two-day drop. A plan sale can still add supply, and it can look ugly beside a customer story. It cannot be read as a decision, made after the SemiAnalysis post, to get out. The motive is not on the form. Mallick has not given one in any source we could open.

The lock-up does not make the Form 4 a contract event. The 10-Q says officers and holders of substantially all of the Class A shares agreed not to sell until the earlier of 6:00 a.m. Eastern Time on the second trading day after earnings for the quarter ending September 30, 2026, or 180 days after the May 14 prospectus, which is November 10, 2026. Exceptions cover some tax withholdings and option exercises. The company estimates up to about 171.1 million shares can come out during that window, including up to about 15.0 million held by Section 16 directors and officers. Mallick’s sale is supply from a plan. It does not revise the customer contract.

Revenue is in a different document. In December 2025 Cerebras signed a master relationship agreement with OpenAI OpCo, LLC. OpenAI is contractually committed to purchase 750 megawatts of inference capacity and related services, over three or four years, extendable by OpenAI to five. OpenAI also holds an option for another 1.25 gigawatts by the end of 2030, which would take the total to 2.0 gigawatts. In the three and six months ended June 30, 2026, Cerebras recognized $56.8 million and $74.4 million under that arrangement, net of $2.5 million and $3.3 million of warrant amortization. Remaining performance obligations were $25.4 billion at June 30, and the 10-Q says a significant amount of that balance is the OpenAI agreement. About 22% of the RPO is expected in the 24 months ending June 30, 2028, about 43% over the following two years, and the rest after that. Delivery dates can move if the customer asks.

The cash around that contract is not the officer’s account. OpenAI funded a secured loan of about $1.0 billion in January 2026. At June 30, $918.2 million was outstanding at 6.0%, repayable in cash or in services. A warrant covers up to 33,445,026 shares at $0.00001, and the 10-Q says 10,033,508 of those shares were issued in a July exercise. None of that moves because a planned sale cleared. The 10-Q already names the revenue risk: OpenAI may buy more compute from competitors, and a serious service-level miss can give OpenAI the right to end part of the agreement or all of it.

What happens next

The quarter ended September 30, 2026 is not on file. The latest 10-Q is the period ended June 30 and the latest 8-K is August 12. There is no newer OpenAI revenue figure than the $56.8 million booked in that quarter.

The first prediction is about that report, not about the stock’s next print. The September quarter will not, by itself, prove the Ultrafast rumor. It will show whether revenue under the OpenAI arrangement is still being recognized. If that line keeps moving off the June base, a single speed tier on Nvidia has not yet appeared as a lost quarter. If management discloses a change to the master agreement, a cut to the 750 megawatts, or a service-level failure, that paragraph is the revenue event. The stock can move before the filing. The filing is still what changes the number.

The second prediction is about supply, and it does not need the Ultrafast story to be true. The lock-up’s early end is tied to the same earnings release, at 6:00 a.m. Eastern Time on the second trading day after it, unless November 10 arrives first. The company has already said that up to about 171.1 million shares can come out during the window. More sales under the June 30 plan can print before then, on the same pattern as the September 29 Form 4. Read those forms as plan executions unless the footnote says otherwise. They add shares for sale. They do not revise the $25.4 billion of remaining performance obligations.

The open customer question is the lane, not the whole contract. Cerebras says it powers GPT-5.6 Sol Ultrafast. Reporters say SemiAnalysis said Nvidia powers GPT-6.1 Sol Ultrafast. Both can still be standing after the next call. What would change the revenue math is a Cerebras disclosure that committed capacity had moved, or an OpenAI disclosure that names the hardware. Neither was in the sources FinanceFeeds could open on October 2. This is not financial advice.

Frequently asked questions

Why is Cerebras stock falling?

CBRS fell from $194.95 on September 29 to $169.52 on October 1, under the $185 IPO price. The attached news is a reported SemiAnalysis post that GPT-6.1 Sol Ultrafast runs on Nvidia, not Cerebras, plus Mallick’s completed sale of 396,000 shares under a June 30 plan. OpenAI and Cerebras had not confirmed the hardware claim.

Is Cerebras below its IPO price?

Yes. The offer was $185.00 and the October 1 close was $169.52, about 8.4% below it and 45.5% below the $311.07 first-day close. The stock has closed under $185 in 20 of 97 sessions. The $160.81 figure is the June 26 intraday low, not the current price.

Did OpenAI move Ultrafast from Cerebras to Nvidia?

Not as a confirmed fact. Cerebras still says it powers GPT-5.6 Sol Ultrafast at up to 750 tokens a second. September 30 reports say SemiAnalysis posted that GPT-6.1 Sol Ultrafast is on Nvidia at a low batch size. FinanceFeeds could not open that post, and neither company had confirmed a switch by October 2.

Did the COO sell, or only file to sell?

Both. The September 29 Form 144 noticed a proposed sale of 396,000 shares worth $77,909,040 on the form. The Form 4 reports that sale as completed the same day, for gross proceeds of $80,798,603.74 at the weighted-average prices. It was a Rule 10b5-1 sale under a plan adopted June 30, 2026.

Does the insider sale change Cerebras revenue?

No. The buyer pays the officer. OpenAI revenue is the master agreement: $56.8 million recognized in the June quarter, against $25.4 billion of remaining performance obligations. A personal sale does not edit those lines. A hardware shift would matter only if it changed how much capacity OpenAI takes.

What is the 52-week range, and what happens at the next report?

Nasdaq’s range is $160.81 to $386.34. The quarter ended September 30, 2026 was not filed as of October 2. The lock-up can end at 6:00 a.m. Eastern Time on the second trading day after that report, or on November 10, 2026, whichever is earlier.