The New Shape of Product Engineering

On August 4th Bending Spoons agreed to buy Airtable for $1.285 billion. Airtable raised $1.4 billion getting there. The company sold for less than the money that went into it, and for about a ninth of the $11.7 billion it was worth in 2021.

A common explanation is that companies were slow to AI and got eaten. Airtable was not slow. It shipped Omni in June 2025, publicly refounded itself as an AI-native company, launched Superagent in January 2026, its first standalone product in thirteen years, then shut Superagent down in April and relaunched the whole thing as Hyperagent on its own domain. Three agent products in ten months, one of them already retired into the next. Whatever killed the valuation, it was not a failure to notice.

Before signing, Airtable reorganized. It moved one product line, Hyperagent, into a separate company called Hyperagent Inc. Bending Spoons bought everything else: the core platform, about $480 million in annual recurring revenue, still growing north of twenty percent.

What the filing shows is that the new product could not be attached to the old company. Somebody did the math and concluded Hyperagent inside Airtable was worth less than Hyperagent beside it. The reason is price.

Marginal cost

For twenty years software cost a lot to build and almost nothing to serve. Getting the first version out took years and salary. Serving the ten-thousandth customer after that cost about the price of bandwidth, ninety cents of gross profit on every incremental dollar forever. That pairing justified 30X revenue multiples, and it made the per-seat subscription an excellent container. If serving is free, the correct strategy is to hang as much as possible off one contract and amortize it across a growing base. Building and distribution were where the money went. Scale was free.

AI inverted both halves. Building got dramatically cheaper. Serving got expensive, and worse than expensive, it got variable, because inference bills per request against a subscription price negotiated before anyone knew what a request would cost.

Canva is the clearest case. It shipped hard into AI: it bought Affinity and made it free, launched Sheets and Code, and its AI features were used at enormous volume. In August it cut its 2026 revenue growth forecast by a third, down to twenty percent, and the stated reason was that demand came in above plan and inference costs broke the model. Canva had already cut its AI serving costs by roughly ninety percent by moving to models it trained itself, to the point where its video model reportedly runs 17X cheaper than a comparable frontier model. That did not save the quarter, because usage grew faster than the savings. Canva then slowed the rollout of a product people liked.

Figma got the same treatment in the same month. Second quarter revenue grew 48 percent to $370 million, earnings beat, full-year guidance went up, and the stock fell about 16.5% after hours. What got priced was the cost line: cost of revenue more than doubled year over year, and non-GAAP gross margin came in at 85 percent against 90 a year earlier. Figma had started metering AI with credits in March, and margin was already recovering, up two and a half points on the quarter. It got marked down anyway, on a beat and a raise, because what the market is now reading is the direction of the cost curve rather than the level of the margin.

You could underprice a seat, overprovision features, give away usage for free and it didn't matter because the incremental cost of serving another user was basically zero. That's not the case anymore.

Jamin Ball, Clouded Judgement, March 2026

The most extreme version is not a subscription company at all. OpenAI built Sora, a credible TikTok competitor, in months. It hit number one on the App Store and passed a million downloads in five days, beating ChatGPT's own record. It then lost about a million dollars a day against roughly two million dollars in lifetime in-app revenue. OpenAI announced the shutdown on March 24th and closed the app on April 26th, taking with it a three-year Disney licensing agreement and the billion dollars of Disney equity investment attached to it, announced in December and never actually paid. Every viral hit made Sora worse off. The consumer platforms I grew up on got their content free, because users shot it, and paid only to store and stream it afterwards. Sora burned compute to manufacture every clip, so the cost sat at creation, and it was paid again every time somebody asked for one more.

The old risk was building something nobody wanted. There is now a symmetrical risk of building something everybody wants, on a contract that cannot absorb them.

Displacement

Cost is not the only thing pushing on these companies.

In August, alongside the forecast cut, The Information reported that Canva executives had noticed users going to ChatGPT to generate designs instead of opening Canva. Serving the product got expensive in the same quarter that a general model started doing the thing the product exists to do.

Anthropic shipped Claude Design in April, pointed at the same category, generating prototypes, decks, mockups and marketing assets from a prompt. It is included in a Claude subscription rather than priced separately, so trying it costs an existing subscriber nothing.

For Canva the two arrive through the same door. The feature that blew out the cost line is design generation, and design generation is what a person now gets from ChatGPT or Claude without opening Canva at all. Canva pays per request to offer the thing its competitors fold into a subscription the customer has already bought.

That leaves Canva pushing against both at once. It has to ship AI features or lose on displacement, and shipping them costs it the margin. Training its own models to cut serving cost by ninety percent was a rational answer to both, and it still lost the quarter.

The insulated companies are insulated for reasons that have nothing to do with model quality. Ramp is a card, a ledger and a set of controls that a finance function is wired into. Stripe is rails. A better model does not make either one easier to leave, because what held the customer was never the generated artifact.

The company says its ChatGPT and Claude integrations have become a meaningful source of user growth. I would still rather not depend on distribution from the product my customers are leaving me for.

Standalone products

Once I started reading for that constraint I found the same move everywhere among companies having a good year.

• Ramp ran an LLM router internally for three years and launched it publicly on August 20th. Not as a Ramp feature. As router.com, on its own domain, with its own signup and its own meter.
• Stripe did not ship stablecoin settlement inside the payments API. It incubated Tempo with Paradigm, which raised $500 million at a $5 billion valuation in October 2025 as a separate company with its own cap table.
• Claude Code started as an internal command-line tool called clide that nobody had commissioned. Boris Cherny joined in September 2024, had a pull request rejected with a note telling him to go try it, pasted the issue in and watched it write the fix. Twenty percent of engineering was using what came out of that on the first day of the November internal release and fifty percent by the fifth. It shipped publicly in February 2025 and reached eight billion dollars of annualized revenue by May 2026, about a sixth of Anthropic's run rate.
• PostHog publishes the rule in its handbook: any member of its Blitzscale team can greenlight a new product even when the others disagree, which is how Session Replay came to be built by an engineer who lost the argument internally and shipped it anyway.

The common thread is a product that can be repriced on its own, without reopening every contract the company already signed.

Agency

Somebody had to start each of those. I think that is the scarce input now, a person who can decide what to build and then execute. Technical skill is the price of entry and most good companies have plenty of it. What is rare is a company that lets the same person do both halves.

Most engineering organizations would say they have no spare capacity. The roadmap is full, and AI has not emptied it. That is the tell. A company that runs product and engineering as separate functions, with requirements handed from one to the other, converts a speed increase into more tickets moving left to right. The companies with visible surplus are the ones that had already stopped working that way. Ramp, Anthropic and PostHog were giving engineers and product people the authority to decide what to build before any of this, so when the cost of building fell, the capacity manifested as leveraged value generation.

A company that runs product and engineering as separate functions, with requirements handed from one to the other, converts a speed increase into more tickets moving left to right.

The published numbers split the labor market by seniority, with developers aged 22 to 25 down about twenty percent from their 2022 peak while their colleagues over 30 in the same roles grew. I read seniority there as a loose proxy for knowing what is worth building, and loose enough that plenty of experienced people are also finding themselves reassigned to work a model could do.

The counterargument

The strongest objection is that I have drawn a causal arrow through a correlation. My winners all sell into the AI build-out rather than being disrupted by it.

OpenAI against Anthropic answers that. Same category, same underlying technology, same customers, comparable capital and talent. Anthropic holds around 54 percent of the enterprise AI coding market against OpenAI's 21, and projects positive cash flow in 2027, while OpenAI projects roughly 14 billion dollars of losses in 2026. What differs is the order of operations. Claude Code was useful to the people who built it before anyone asked whether it could be sold, and it went on sale only after the company's own engineers had proved they would not work without it. The problem was solved, then priced, then scaled. Sora ran the other way. It was built because it could be, launched at an audience nobody had confirmed would pay, on a cost base nobody had checked against that audience, with no owner whose job was to close the gap between the two. It was dead in seven months.

Meta's generational blunder

Meta is the case I find hardest to write about calmly. Meta's business is fine. Revenue grew 33 percent year over year in the first quarter of 2026 and 28 percent in the second. This is not a revenue problem.

That cash machine is a lagging asset. It was built between roughly 2009 and 2016, by News Feed and the mobile transition and the Instagram acquisition, inside an organization that rewarded engineers for moving without asking. Michael Novati, who was there for most of it, describes rewriting Facebook's org chart tool as an intern without permission and having it land well, in a company where breaking something while moving fast did not get an engineer fired. That company treated internal tools as real products, which is most of what I have argued above, and Meta got there first.

The 2026 spending is where it gets harder to defend. Reality Labs lost more than 80 billion dollars cumulatively with nothing on a path to profitability, capital expenditure is guided to $130 to $145 billion, free cash flow is down 91 percent, and about 6,500 engineers and product managers were moved into an applied AI unit to generate training data, many notified by email. That is the same surplus capacity, pointed at a task that produces no product and no compounding asset. Andrew Bosworth, Meta's own CTO, wrote in June that the company had done an atrocious job explaining what the unit was for, and told staff that morale was close to the worst he had seen in twenty years. People inside the unit called themselves draftees. Meta responded with retainer equity grants, four hundred thousand to a million dollars over three years for senior engineers who resigned, and larger ones aimed at the specific people Anthropic was hiring.

That is the same surplus capacity, pointed at a task that produces no product and no compounding asset.

Meta is converting a decade of accumulated talent advantage into a decade of accumulated capital expenditure. The failure will not show up in a quarterly result. It shows up as an absence, as the product that does not get built in 2027 by the person who left in 2026.

Every company on the losing side of this had good engineers and knew about AI in 2023. What they did not have was anywhere to put a new product except inside the old one. That is a decision about org structure that gets made long before it gets tested, and most of the companies that will fail this test in 2028 have already failed it.