Amazon is in talks to buy Decart, a San Francisco startup whose software lets developers move workloads between different AI chips, in a deal that could be worth around $7 billion. A few items further down the same Wall Street Journal newsletter, Apollo's chief economist Torsten Slok concludes that "the productivity payoff from AI remains a forecast rather than an observation." Put those two items together and the picture of the AI deal market changes.
The current wave of AI acquisitions is being driven by a jammed exit market and by big buyers scrambling for chips, compute and people. It is not being driven by evidence that AI makes companies more productive. Acquirers should price these deals as strategic options, with a real premium and a real chance of expiring worthless, and boards should stop reading the deal flow as proof that AI pays.
The occasion is Steven Rosenbush's Morning Download for October 12, which pitches M&A as "the alternative exit" for AI startups that cannot go public. Alongside Decart, it notes that Nvidia is in early talks, reported by the FT, to deepen its $800 million investment in open-weight model developer Reflection AI, possibly through an acqui-hire or a chips-and-compute deal. It also reports that quantum computing company IonQ has agreed an all-stock purchase of cybersecurity firm Elisity at up to $1.3 billion, after a $1.8 billion takeover of chipmaker SkyWater Technology earlier this year.
The buyers are paying for access
Amazon wants a credible answer to Nvidia for its Trainium chips, and Decart, which got early access to Trainium in 2025, makes moving between chips easier for developers. Decart's world models, which simulate environments where robots and autonomous vehicles can learn, suit a company that operates in the physical world at scale. Nvidia's options for Reflection amount to a talent grab or a way to lock in demand for its own hardware. IonQ is assembling a portfolio with its own shares as currency. None of these rationales depends on an enterprise customer somewhere getting more done in fewer hours.
That is defensible as strategy. Securing a supply chain, a team or a platform position is a legitimate reason to spend billions. The trouble starts downstream, when the headline price of an AI deal becomes a reference point for every corporate development team and investment committee, and slowly gets reinterpreted as a market verdict on what AI is worth to the businesses that use it.
The selling side makes the same point from the other direction. PitchBook's Q3 2026 Venture Monitor counts just 18 IPOs in the quarter and 992 active unicorns with no clear path to liquidity, and it records Airtable selling for $1.3 billion and Miro for $1.4 billion, both at discounts of nearly 90 percent to their prior valuations. The M&A door is open, but for many companies it is the exit of the cornered. When one market produces a $7 billion bid for a chip-portability startup and a 90 percent haircut for well-known software firms, the dividing line is scarcity, and nobody in either transaction was pricing measured productivity.
The productivity case is still a forecast
If deal prices were tracking demonstrated returns, the macro data would show something by now. It mostly does not. The St. Louis Fed tagged productivity statements on corporate earnings calls and found that roughly 95 percent of AI-related ones referred to future gains, against about three-quarters for other productivity talk. The same analysis cites San Francisco Fed figures showing utilization-adjusted total factor productivity grew only 0.07 percent over the four quarters ending in the first quarter of 2026. Executives are describing a payoff they expect; the economy has yet to record one.
The newsletter offers a plausible reason, drawn from a New York Times DealBook research roundup: workers spend much of the time AI saves checking and fixing its output. For acquirers, that has a direct consequence. If the binding constraint is how much expert humans can review, buying more model capability does not loosen it, and an acquired AI business inherits the same bottleneck its customers face.
The strongest case against our view goes like this. Capability gaps in chips, models and specialist talent close slowly, and companies that secure them early lock in advantages that late movers cannot buy back at any price. Productivity statistics lag adoption; the San Francisco Fed notes that its regime-switching model is reacting much as it would have in the mid-1990s, which it calls reason for "cautious optimism." On that reading, waiting for proof means arriving after the assets are gone.
We accept most of that, and it strengthens our argument. A bet on scarce capability whose payoff may arrive in three years or ten is the textbook definition of an option. The mistake is booking it as though the payoff had already been observed. Options carry premiums, conditions and expiry dates. An acquirer who pays $7 billion for strategic position has made a defensible call; an acquirer who justifies $7 billion with productivity gains the data cannot yet see has made a forecast and labelled it a fact.
Price the option, then decide what to pay
Treating these deals as options changes what an investment committee should demand. Instead of modelling synergies from efficiency gains that firms themselves only forecast, it should be able to answer four questions:
- What specific scarcity does this buy (a chip ecosystem, a model, a team), and how long before a competitor can get it another way?
- What is the asset worth if the productivity payoff arrives five years late, and can we carry it that long?
- How much expert human review will the acquired capability demand once it is inside our business, and who supplies it?
- If the strategic rationale collapses, what remains: talent that stays, technology that transfers, or a write-down?
Committees are most likely to skip the third question, because acqui-hires and capability buys tend to be modelled as additions to capacity. If the DealBook finding holds, every AI capability brought in also loads review work onto the experts the company already has, and those people are among the scarcest resources it owns. A deal model that ignores that cost will flatter the purchase price by exactly the amount the organisation later spends on supervision.
An acquirer who justifies $7 billion with productivity gains the data cannot yet see has made a forecast and labelled it a fact.
Stock-financed buyers deserve extra scrutiny here. When the currency is your own shares, as in IonQ's Elisity deal, the cash discipline that normally forces a hard look at returns is weaker, and an option can be overpaid for without anyone noticing until the share price does.
The exit logjam will keep the M&A door open; PitchBook reckons the backlog of companies that should have gone public could fill more than two full years of IPOs. That means more deals, bigger headline numbers and more temptation to read them as a scoreboard for AI's value, a temptation boards should refuse. The buyers in this newsletter are paying for chips, compute and people they cannot build fast enough, and they may well be right to. The productivity case is a separate claim with its own evidence, and so far that evidence is mostly earnings-call promises. A company that confuses the two will overpay twice: once for the asset, and again when the returns it booked at signing fail to arrive.
