Amazon shifts AI focus to business impact, not just raw usage, in new developer guidelines.
The company warns that AI metrics can create bad incentives, including “tokenmaxxing.”
Managers are told to examine teams at both the highest and lowest levels of AI adoption.
Amazon just issued new guidelines for software engineers in the AI era: Prioritize business outcomes over raw usage.
Terk Terkowitz, Amazon's vice president and distinguished engineer, said in a memo published Tuesday that AI is changing software development more quickly and fundamentally than earlier shifts such as cloud computing. But engineers should be judged by their impact on customers and the business, he wrote, not simply how much they use AI.
"We all need to embrace this change and take the time to learn, or we risk being left behind," Terkowitz said.
AI coding tools and agents are upending the software industry and changing what it means to be a coder. Amazon calls this the "era of the Renaissance Developer," where human judgement and machine intelligence combine to create new products and services faster and more efficiently.
"This is a time of change for our industry, and a time of change for what it means to be a software engineer," Terkowitz wrote in the memo.
The updated guidelines reflect Amazon's growing effort to make AI adoption part of how it assesses engineers and engineering managers. The company has tracked measures including AI usage, token consumption, and lines of code internally, and used leaderboards to rank teams by some of those metrics.
But the memo warns that using such measures as targets can produce bad outcomes, including "tokenmaxxing" rather than better software or customer experiences.
"Using these metrics could, if used as a target instead of a measure, drive negative outcomes," Terkowitz wrote in the memo.
Amazon has been experimenting with ways to measure and encourage AI use across its engineering organization.
Earlier this year, the company was tracking engineers's use of internal AI tools and some teams used leaderboards tied to measures such as token consumption. The programs created friction among some employees, who worried that pressure to use AI could reward wasteful activity rather than productive work.
The new guidance offers Amazon's clearest public acknowledgment of that risk. Rather than lay out a detailed AI playbook that could quickly become obsolete, Amazon said its revised developer guidelines focus on long-term impact on customers and the business.
Still, the company said AI should be "at the center of software engineering practice" and urged engineers to incorporate the technology into their daily work.
Leaders are expected to use AI tools on production work themselves, understand where they help and where they fall short, and explain that judgment to their teams, according to the memo.
Terkowitz said Amazon has found that metrics such as token use, AI-tool adoption, and code reviews can correlate with the outcomes it values. But he said the measures must be used carefully. More token use can raise costs, while a focus on lines of code can create unnecessary complexity.
He also recommended that managers pay particular attention to teams at both ends of the AI adoption spectrum: those in the bottom 10%, to understand why they are not using AI effectively, and those in the top 10%, to determine whether heavier use is translating into faster delivery.
Terkowitz said Amazon's most AI-forward "frontier teams" have shown stronger productivity and delivery. He also said teams and individuals that are not adopting AI or increasing output metrics are falling behind.
But his memo left room for engineers whose work may not yet benefit from current AI tools. In those cases, he said, employees should document why and work toward a solution by sharing their findings with peers and managers.
"An engineer who can explain where AI does not yet help their work is demonstrating the judgement we are asking for," the memo said.
Anthropic is touting its ability to use different AI chips in its initial public offering filing.
Analysts say a multichip strategy could weaken Nvidia's pricing power.
Switching chips is hard and expensive, putting it out of reach for smaller players.
Anthropic is hedging its bets by using multiple chips to handle its AI work — a strategy that could eventually erode Nvidia's dominance.
The AI lab says it manages compute "as a single, fungible pool from multiple strategic partners" — with software that can direct its work to different chips depending on where they're most productive, according to Anthropic's S-1 filing, which came out Monday as the company prepares to go public.
Anthropic says it uses hardware from four chip families: Amazon Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs.
That flexibility, which Anthropic said "took years and significant investment to build," could make major AI companies less dependent on Nvidia over time and give them more negotiating leverage, analysts said.
"The real exposure could be Nvidia's pricing power," said Patrick Moorhead, founder and CEO of Moor Insights & Strategy, "because a customer that can shift work across four platforms negotiates very differently than one that can't."
Ashish Nadkarni, an analyst at IDC, added that a multichip approach also makes Anthropic less vulnerable to fluctuations in Nvidia's chip supply.
Breaking Nvidia's grip isn't easy
Although Anthropic's chip mix isn't new, the company is now pitching it as a core advantage to investors, Moorhead said.
The threat has its limits. Switching between chips is still difficult — especially when it comes to training AI models as opposed to running them, Moorhead said — and only the biggest AI companies have the resources to do it.
Anthropic also continues to spend billions on Nvidia compute, Moorhead said.
Gil Luria, head of technology research at D.A. Davidson, said that while other major companies like OpenAI likely take a multichip approach, it may not be worth the effort and investment for smaller players.
Brad Gastwirth, global head of research and market intelligence at Circular Technology, said Anthropic's strategy doesn't represent a major near-term threat. Having more chip options doesn't erase Nvidia's advantages in performance, networking, and software, he said.
There's a broader risk for Nvidia if AI compute becomes more of a commodity.
Citadel CEO Ken Griffin has warned that if supply overtakes demand, it could drive down rental prices.
"GPUs, like airlines, have a very low marginal cost to run," he told Bloomberg last week. "The moment you have overcapacity, pricing goes into freefall."
Nvidia did not immediately respond to a request for comment from Business Insider.
Have a tip? Contact this reporter via email at gweiss@businessinsider.com or Signal at @geoffweiss.25. Use a personal email address, a non-work WiFi network, and a non-work device; here's our guide to sharing information securely.
Two associates of Sam Bankman-Fried invested in Anthropic. After they were sentenced for fraud, their shares entered a legal black hole.
ANGELA WEIS/Getty Images; Tyler Le/BI
Sam Bankman-Fried's Anthropic stake was liquidated in bankruptcy court two years ago.
The feds seized shares belonging to two co-conspirators, Caroline Ellison and Nishad Singh.
The government quietly sold those Anthropic shares last year. It's unclear where the money will go.
A little over a year ago, the federal government acquired a stake in Anthropic without paying a penny.
The shares were seized from associates of the crypto fraudster Sam Bankman-Fried. From there, they disappeared into a legal black hole.
As Anthropic rockets toward what could be the biggest IPO in history, those shares could now be worth billions of dollars.
Bankman-Fried's own Anthropic stake was liquidated in the bankruptcy of his failed cryptocurrency exchange. The Anthropic equity owned by Caroline Ellison and Nishad Singh, two associates who invested alongside him, took a different path. The government seized those shares and sold them to existing Anthropic shareholders, according to a person familiar with the sale.
It's still unclear whether the proceeds will be paid out to victims of the $11 billion fraud case, or if the feds will just keep the money.
It also remains a mystery which investors bought Singh's and Ellison's stock and how much they paid, but Anthropic shares have been on such a dizzying climb that it's likely those investors made a killing on them.
Bankman-Fried was convicted in 2023 on fraud and money laundering charges and sentenced to 25 years in prison after the collapse of his cryptocurrency empire. Prosecutors said he used his crypto hedge fund, Alameda Research, to siphon billions of dollars from customers of FTX, his crypto exchange.
The story of the crypto criminals' Anthropic shares is a tale with extraordinary elements, including one of the biggest financial frauds in modern history, a messy bankruptcy docket, and a company that experienced one of the fastest appreciations in business history as the world suddenly realized AI's existential importance.
An intelligent investment
Bankman-Fried, Singh, and Ellison each invested in Anthropic's 2022 Series B funding round. Bankman-Fried bought $500 million worth, which according to court records represented 13.56% of Anthropic at the time. Singh acquired $40 million in shares, and Ellison acquired $10 million in shares, court records reviewed by Business Insider show.
The Anthropic shares purchased by Singh and Ellison could together be worth between $4.17 billion and $5.03 billion today, based on the $965 billion valuation the company announced this May, according to Olav Sorenson, who teaches venture capital strategy at UCLA. Harrison Rolfes, an analyst at PitchBook, put the combined figure at $2.62 billion. If Anthropic went public at a $2 trillion valuation, the shares would be worth about $5.44 billion, Rolfes said.
Among Sam Bankman-Fried's many investments was a sizable stake in Anthropic, the AI company that has since soared in value.
Bloomberg/Getty Images
Four of Bankman-Fried's close friends and executives at his companies pleaded guilty as co-conspirators. Three of them, including Singh and Ellison, testified against him.
After the collapse of FTX, Bankman-Fried's Anthropic shares — which were owned by Clifton Bay, an entity affiliated with Alameda Research — ended up in bankruptcy court along with the other dredged-up remains of his companies. His Anthropic shares were liquidated to pay FTX's creditors.
The estate in 2024 sold Bankman-Fried's Anthropic shares to a few dozen buyers for a total of $1.3 billion, more than double what he paid. The largest stake went to an entity affiliated with a United Arab Emirates sovereign wealth fund, bankruptcy court filings show.
The feds take a stake in Anthropic
As part of their sentences, a judge required Singh and Ellison to forfeit their Anthropic shares, which prosecutors said could be considered proceeds of their crimes.
At Singh's sentencing hearing, one of his attorneys, Andrew Goldstein, said Singh purchased his shares before participating in the criminal conspiracy and "he actually may have had a legitimate legal claim" to the shares but agreed to give them up as part of his plea agreement "because it was the right thing to do."
Reached for comment for this story, Goldstein told Business Insider that Singh hopes the government is able to quickly distribute proceeds of the sale to FTX victims. An attorney for Ellison declined to comment.
Nishad Singh was required to give up his shares of Anthropic as part of his sentence.
Bloomberg/Getty Images
A federal judge ordered Ellison's and Singh's Anthropic shares to be transferred to the federal government, which took ownership of Ellison's shares in February 2025 and Singh's in April of that year, according to previously unreported court records.
Ordinarily, victims of crimes are compensated through a restitution process, which is overseen by courts. But the number of potential FTX victims could be in the millions, prosecutors said in court filings. As a result, the judge ruled, victim compensation would be handled through a process called remission, which is overseen by the Justice Department.
During Ellison's sentencing hearing, Justice Department prosecutors told the judge that the DOJ would either set up its own claims administration process to compensate victims, or work with the FTX bankruptcy process to identify victims and provide forfeited funds to them.
At the time, FTX's bankruptcy process was in full swing. A Delaware court appointed Sullivan & Cromwell, the elite Wall Street law firm, to untangle the company's assets, figure out who was owed money, and pay them.
The creditors in the FTX bankruptcy, prosecutors noted, largely overlapped with FTX's victims. They were generally depositors, lenders, and investors who were defrauded by Bankman-Fried and other executives. Prosecutors said the Justice Department could work with the bankruptcy estate's lawyers to get money back to them, as it had in previous large-scale financial frauds such as Bernie Madoff's Ponzi scheme.
There was an unusual twist that set the FTX bankruptcy apart. Bankman-Fried's investments — especially the Anthropic shares, along with some cryptocurrency — had grown substantially in value since FTX's collapse. Earlier in 2024, the Sullivan & Cromwell lawyers who'd taken over FTX projected that all the creditors would be repaid in full, with interest.
The Marshals take control
While prosecutors said they intended to use Singh's and Ellison's Anthropic shares for remission, the Justice Department could technically do whatever it wanted with them, according to Duncan Levin, a white-collar defense attorney who teaches a course on forfeiture at Harvard Law School.
"It's a very opaque process," he said. "It's completely at the discretion, by law, of the attorney general of the United States."
Typically, when the feds seize shares of private companies through criminal asset forfeiture, they send the shares to the US Marshals Service Complex Assets Unit for liquidation. The unit tries to value the shares as any other investor would, said Michael Bachner, a white-collar criminal and securities litigation attorney.
"They may look to: What would an institutional purchaser pay for these securities?" Bachner said. "Are there funds that are valuing the securities? Is there a secondary market already out there?"
For the sale of the Anthropic shares, timing was crucial. The Marshals Service is supposed to preserve as much value as possible, Bachner said. By the time the feds got hold of the shares, Anthropic had become an economy-shifting AI giant. In its March 2025 Series E fundraising round, Anthropic was worth $61.5 billion. By its Series G round at the start of 2026, it was worth $380 billion.
At the same time, each fundraising round diluted the Series B shares. The FTX estate told the bankruptcy court that Bankman-Fried's shares, which in 2022 represented 13.56% of Anthropic, represented 7.84% of the AI company in January 2024.
Sunil Kavuri, a victim of Sam Bankman-Fried who has advocated for FTX creditors, said it would be "diabolical" for the government to hold onto the proceeds.
David Dee Delgado/Getty Images
The Marshals Service sold Singh's and Ellison's shares to existing Anthropic investors sometime last year, according to the person with knowledge of the sale.
It's not clear when exactly the agency sold the shares, to which investors, how those investors were chosen, at what price the shares were sold, or how much money the government made in the sale. Depending on when they were sold in 2025, the combined shares could have been worth between $300 million and $1.1 billion, according to Sorenson, the UCLA professor. Rolfes, the PitchBook analyst, estimated a range between $250 million and $630 million, depending on the timing.
The Marshals Service declined to comment. A representative for the Justice Department said information about asset sales and victim compensation is confidential.
The revenue from the sale of Singh's and Ellison's Anthropic shares doesn't appear to have been transferred to the FTX estate as of the end of June this year, according to bankruptcy court filings from the estate, which continues to compensate victims and creditors.
The FTX estate received $638 million last year from assets seized by the Justice Department, according to the estate's annual report for 2025. It's not clear from the filings where those funds came from. The estate expects to receive about $400 million more from the government at some point in the future, according to the annual report. That would include proceeds from cryptocurrency and other investments made by Bankman-Fried.
One of Bankman-Fried's victims, Sunil Kavuri, told Business Insider that the Justice Department should use the proceeds of Ellison's and Singh's Anthropic shares to compensate victims. It would be "diabolical" for the government to hold onto the proceeds, he said.
Victims haven't actually been made whole, he said, because the bankruptcy court calculated the losses of FTX depositors using the time of FTX's bankruptcy declaration, when crypto prices were at a low ebb. Kavuri previously argued in bankruptcy court that the FTX estate should have held on to its Anthropic shares for longer to take advantage of the company's swift growth.
Representatives for the FTX estate declined to comment. The Justice Department spokesperson said the matter was ongoing, and that the DOJ "prioritizes victim compensation from forfeiture and takes all steps to ensure forfeited funds are provided to victims."
The government could end up just keeping the money, Bachner said.
"They've invested millions and millions and millions of dollars in prosecuting Bankman-Fried. And they want to get at least reimbursed for their costs, so sometimes they'll do that," he said. "It's really a unilateral government decision."
Jack Newsham and Katie Roof contributed reporting for this story.
Correction: August 31, 2026 — An earlier version of this story misstated the number of co-conspirators who testified against Sam Bankman-Fried. It was three.
Update: August 31, 2026 — This story has been updated to reflect Singh's role in the FTX fraud case.
Correction: October 27, 2026 — FTX bankruptcy court filings don't make clear where the Justice Department obtained the $638 million it transferred to the estate. A previous version of this story mischaracterized the filings.
Machenta cofounders Edward Johns and Vitalis Vosylius
Oscar May
Machenta raised $40 million in seed funding led by Creandum and Union Square Ventures.
It says robots can learn tasks from one demo using a model trained entirely in simulation.
The startup remains focused on research as it takes on multibillion-dollar rivals.
A London-based robotics startup is betting it can teach robots new tricks without collecting the mountains of real-world data its much bigger rivals are chasing.
Machenta, founded by researchers from Imperial College London, has raised a $40 million seed round to build an AI model that can teach robots new physical tasks after seeing them performed just once, the company exclusively told Business Insider. Venture capital firms Creandum and Union Square Ventures co-led the round. Machenta declined to disclose its valuation.
While rivals building general-purpose AI models for robots are scrambling to amass real-world data, Machenta is taking a contrarian approach. Edward Johns, the startup's CEO and cofounder, and an Imperial professor, said the company trains its model in computer-generated environments that mimic how objects move and interact, without using a "single pixel of real-world data."
Once trained, the model can direct a robot to perform a new task, such as tidying a table or putting objects into a box, after watching it done just once, without further training, Johns said. In a video shared with Business Insider, three robots pack and close a box on an assembly line, a task Johns said they learned this way.
The startup has 11 employees and remains focused on research, with no near-term plans to commercialize its technology. Its model can learn some tasks immediately, but more precise tasks and complex movements remain challenging.
Machenta faces rivals with multibillion-dollar valuations in one of robotics' most competitive and well-funded fields. Its California-based competitors include Physical Intelligence, valued at roughly $11 billion, and FieldAI, which is raising funding at a $10 billion valuation, Business Insider reported this month.
Another rival, Skild AI, which is valued at over $14 billion, is already selling technology with similar capabilities. In August, Skild unveiled a model that it said could perform previously unseen tasks lasting up to 10 minutes after a single video demonstration. Last month, it said that it had reached $100 million in annual recurring revenue from more than 60 customers.
From research lab to startup
Six years ago, Johns began investigating whether robots could learn new physical tasks as quickly as humans at his Imperial College lab.
In the lab's early experiments, researchers physically guided a robot through a task, such as picking up a spatula and scooping an object. The robot then spent about 30 minutes collecting additional data about the object and its surroundings before it could perform the task on its own.
Johns and former Ph.D. student Vitalis Vosylius spent the next few years working to eliminate that extra step, founding Machenta late last year after seeing early signs that their model could perform new tasks immediately after a demonstration.
Machenta's team
Oscar May
Companies are increasingly buying and collecting data to teach robots how to interact with the real world. But Johns believes simply collecting more data isn't the answer. It also remains unclear how much training data will be needed to make general-purpose robots a reality. Johns said that Machenta is taking on the harder challenge of designing algorithms that can generate synthetic data useful for teaching robots physical tasks.
Machenta plans to use the funding to expand its team and train its model with more computing power and simulated data, hoping to teach robots to perform more precise and complex tasks.
Have a tip? Contact Rya Jetha via email at rjetha@businessinsider.com or Signal at rjetha.07. Use a personal email address and a non-work device; here's our guide to sharing information securely.
Panos Panay, head of Amazon's Devices and Services team
CHARLY TRIBALLEAU/AFP via Getty Images
Amazon launched a series of new devices this month, including Alexa tablets and Bee AI wearables.
Rising memory costs forced "difficult trade-offs" on Amazon's device prices, sales, and profit.
Amazon is seeking more Alexa revenue from enterprise sales, licensing, and ads.
Amazon launched a flurry of new consumer devices this month. Behind the scenes, rising memory costs are adding pressure to its already money-losing hardware business.
An internal planning document obtained by Business Insider shows Amazon forecasting a $612 million gross loss this year for its first-party devices business, compared with a leadership target of a $438 million loss.
The document said rising memory costs were "putting pressure" on the business's profitability, forcing "difficult trade-offs" among sales volumes, average selling prices, and gross profit.
The projected $612 million loss was at the gross profit level, meaning it did not account for all operating expenses, such as R&D, sales and marketing, and administrative costs.
The figures offer a rare glimpse into the pressure on Amazon's hardware business as it absorbs rising component costs. Other tech hardware companies, including industry leader Apple, face similar pressure from surging memory chip prices.
An Amazon spokesperson declined to comment.
A new wave of AI devices
Despite the pressure, Amazon continues to expand its consumer hardware lineup. Even if the company loses money on device sales, it can make money in other ways, such as subscription services, as more customers use its gadgets.
In October alone, the company introduced new Kindles, Alexa-powered tablets, Ring smart locks, and AI wearables under its Bee brand. The August planning document identified Bee and new smart glasses as strategic products still in an investment phase.
Amazon has historically used low-cost hardware to pull customers into a broader ecosystem of shopping and digital content. Rising component costs could make that strategy more expensive, forcing the company to raise prices.
The clearest shift is in Amazon's tablets. Earlier this month, the company phased out Fire tablets after 15 years and replaced the budget-friendly lineup with pricier Alexa Tablets powered by Alexa+. The new devices cost between $229.99 and $499.99, a notable increase from previous generations.
Not just Amazon
The planning document was prepared in August, the same month Amazon raised prices on products including Echo speakers, Kindle e-readers, and Fire TV devices.
The pressure extends beyond Amazon. Apple has raised prices on some Macs and iPads, while Microsoft and Sony have increased console prices. Dell, HP, and Lenovo also face higher memory costs.
The AI boom is helping drive the squeeze. Huge data center investments have increased demand for memory chips, while manufacturers have devoted more capacity to high-bandwidth memory used in AI servers. That has tightened supplies and raised prices for components used across consumer electronics.
Amazon is experiencing both sides of the boom. It expects roughly $220 billion in capital expenditures this year, largely to expand AWS data centers for AI demand, while higher component costs squeeze its consumer hardware business.
The rising cost of Alexa+
The August planning materials show Amazon reassessing both Alexa+ revenue and operating costs. They called for updated forecasts, expansion through enterprise sales and advertising, and more detailed AWS cost modeling.
Amazon is now confronting the cost of operating Alexa+ at scale. Business Insider reported in July that internal forecasts put Alexa+'s AWS cloud costs at roughly $1.7 billion in 2026, nearly triple the prior year's figure. Amazon was working to lower those costs by relying less on Anthropic's AI models and routing more requests to its own, cheaper systems.
A separate project called Moonraker, intended to give Alexa more advanced AI-agent capabilities, was projected to require more than $100 million in GPU spending this year. Amazon internally considered delaying or scaling back the initiative as one option for easing cost pressure, Business Insider previously reported.
Intel CEO Lip-Bu Tan, Hyperbolic founder Jasper Zhang, and Walden Catalyst partner Shankar Chandran.
Luke Dietz
Infrastructure startup Hyperbolic raised a $30 million round led by Intel CEO Lip-Bu Tan's VC firm.
The startup connects companies with neoclouds and makes their AI chips easier to use.
Hyperbolic also wants to let customers resell unused computing power.
Intel CEO Lip-Bu Tan is backing an infrastructure startup that has raised $30 million to help companies rent compute across hundreds of different AI cloud providers.
Hyperbolic, an AI chip marketplace and software startup, has raised $50 million to date.
Walden Catalyst, the venture capital firm founded by Tan and former Samsung president Young Sohn, led the Series A-1 funding round. Dell Technologies Capital, General Catalyst, Vista Equity Partners, and Cambium Capital also participated.
As companies scramble for computing power, one challenge is that AI cloud providers, or neoclouds, use a mix of different chips, networking, and software. Hyperbolic helps customers find compute resources from different providers and creates software that makes it easier to use them all, rather than dealing with each one separately.
"It's been a really fragmented market that makes the customer experience really painful," said 31-year-old founder and CEO Jasper Zhang.
Hyperbolic also wants to let customers resell compute that they're paying for but no longer need.
The four-year-old startup works closely with about 30 neoclouds, including WhiteFiber and Vector Core Compute. It says it has roughly $200 million worth of compute connected to its software annually.
Hyperbolic previously raised a $12 million Series A round in 2024, led by Variant and Polychain Capital, following a $7 million seed round earlier that year. The company expects to have 50 employees by the end of the year and will use its funding to expand software tools for managing AI chips across different neoclouds.
Zhang said the company is aiming for $150 million in annual recurring revenue by the end of the year and is already breaking even. More than 250,000 developers, including teams at Modal, Higgsfield, and Daytona, use Hyperbolic.
Before founding the company, Zhang worked as a quantitative researcher at Citadel Securities. As Wall Street explores ways to financialize AI chips, he said his startup's plans to let companies resell unused compute could help pave the way for a futures market.
Here's a look at the pitch deck Hyperbolic used to raise its $30 million Series A-1. Slides have been redacted so that the deck can be shared publicly.
Posters on a wall at Anthropic's developer conference in San Francisco.
Alistair Barr/Business Insider
Anthropic shows strong profit potential, with revenue growing way faster than expenses.
Anthropic's revenue per employee exceeds $6.6 million, outpacing estimates for Meta and Google.
Anthropic has a surprisingly diverse customer base and huge cash reserves.
For years, one of the biggest unanswered questions about the artificial intelligence boom has been surprisingly simple: Can the companies building the world's most powerful AI models actually make money?
After spending hours poring over Anthropic's IPO filing on Monday, we finally have an answer: Yes. And potentially lots of it.
A couple of Chinese AI labs went public earlier, revealing appalling financial results that sowed doubt about the industry's economics. Now Anthropic has provided a much more compelling demonstration of how this business can work at enormous scale.
The numbers are remarkable. Anthropic generated $11.6 billion in revenue during the second quarter of 2026, up 1,368% from a year earlier. During that same period, spending on computing and infrastructure increased 300%, while other expenses rose 741%
So, the company is bringing in money much faster than its most importantcosts are rising.
For every dollar Anthropic spent on computing in the second quarter, it generated $2.08 in revenue, up from just 40 cents in the first quarter of 2025. That's a more than fivefold improvement over five quarters.
The result: Anthropic recorded $974 million in quarterly operating profit, representing an8% margin. Its adjusted operating profit margin reached 19%, compared with -129% a year earlier.
"Frontier-lab unit economics can turn positive at scale," Mark Mahaney, a top tech analyst at Evercore ISI, wrote in a research note on Monday.
Better than Meta and Google
There's more evidence that Anthropic has built an unusually efficient business.
Over the year ending September 30, revenue increased roughly 1,200%, while its employee count grew about 170%. Revenue per employee jumped from $1.4 million to more than $6.6 million.
For comparison, Mahaney estimated that Meta generates about $3.2 million in annual revenue per employee, while Google generates about $2.5 million.
That comparison isn't perfect. Anthropic relies heavily on cloud providers, whose employees handle much of the computing infrastructure and distribution work. The improvement is still striking.
Customer surprise
Another surprise: Anthropic doesn't depend as much on a handful of enormous customers, as some investors had feared.
The company had more than 11,900 customers generating at least $100,000 in annual revenue as of September, compared with over 950 a year earlier. More than 1,900 customers spend at least $1 million annually.
No single customer accounted for more than 10% of revenue during the first nine months of 2026. Its five largest customers represented about 21% of third-quarter revenue, down from roughly 36% in the first quarter.
"Concentration may be better than feared," Mahaney wrote.
Cash cushion
Anthropic has a formidable financial cushion. Its June 30 balance sheet showed more than $130 billion in cash and short-term investments.
And Mahaney estimated that the company has burned through just $9 billion since it started more than five years ago. That's a long runway.
Now, don't ask me whether all this makes Anthropic worth $2 trillion, $1.5 trillion, $1 trillion, or considerably less. If I knew that, I'd be making a lot more money as an investor. And even professional investors don't know.
An $800 billion caveat
Some enormous caveats: Much of Anthropic's revenue comes from customers paying for what they use, rather than signing long-term contracts. Meanwhile, some of its computing agreements require minimum payments whether or not it needs all the capacity.
Mahaney estimates that Anthropic has entered into nearly $800 billion in computing arrangements, although only about $266 billion was fixed or guaranteed as of June.
That's still a staggering amount.
"The commitments assume revenue compounding that is not contracted," Mahaney wrote.
Translation: While the company has committed to spending hundreds of billions of dollars on AI computing infrastructure, it hasn't secured the future revenue needed to pay for all of it.
Importantly, most of the potential $800 billion exposure isn't fixed. Many arrangements depend on capacity being delivered, can be adjusted, or include cancellation provisions.
In-house compute and capex
Another wrinkle: Anthropic warned Monday that its spending will increase as it shifts toward leasing AI chips and operating more of its own computing infrastructure, rather than relying entirely on Amazon, Google, and other cloud providers.
That could make future cash flow considerably less attractive, even if operating profits continue to improve.
Still, the central takeaway from Anthropic's IPO filing is difficult to ignore.
For years, investors have wondered whether selling AI could become a genuinely profitable business, rather than an extraordinarily expensive scientific experiment.
Anthropic has now shown that it can. The next question is whether it can keep doing so as the bills get much, much bigger.
Login
