a16z Report Reveals AI ROI Measurement Crisis at Scale

Let's talk about the a16z State of Markets report, because this is one of those documents that quietly reframes the entire AI conversation — from "what can the models do" to "where is all this money actually going, and is anyone checking." **Technical Deep Dive** The core technical shift a16z is documenting isn't about model architecture at all — it's about where capital and compute are physically landing. Alphabet, Amazon, Meta, Microsoft, and Oracle spent $416 billion on capital investment in 2025.

Analyst estimates now put 2026 spending at $777 billion, with a forecast of $1.1 trillion for 2027. That money isn't buying chatbots — it's buying chips, power substations, cooling systems, and networking gear.

One detail that stood out to us: NVIDIA's older A100 chip rental prices have stayed flat or even risen since the start of the year, despite newer silicon flooding the market. That's not supposed to happen in a normal hardware cycle — older generations are supposed to depreciate. Instead, demand for raw compute, any compute, is so intense that even "obsolete" chips are holding value.

Layer in Joanna's earlier point about Google's Argon system reportedly freeing up a petabyte of datacenter memory, and you start to see the full picture: companies are simultaneously building new capacity and squeezing more out of existing infrastructure, because neither alone is keeping pace with demand. **Financial Analysis** Here's where it gets uncomfortable. Big tech's own profits funded a lot of this buildout initially, but now borrowing is stepping in — Bloomberg reported Broadcom alone is amassing $60 billion to fund chips for Anthropic.

That's debt-financed AI infrastructure, which changes the risk calculus considerably. Meanwhile, cloud providers' revenue backlogs — customer commitments not yet recognized as revenue — have more than doubled year over year, which sounds bullish until you remember backlog isn't cash in hand, it's a promise. And on the demand side, only about 2% of U.

S. households paid for any AI service in April. That's the real gap in this story: massive, debt-amplified infrastructure spending chasing a consumer market that's barely monetizing yet.

Software companies are feeling a parallel squeeze — about 75% of public software firms are profitable, but only around 30% are growing 20% or more annually, and in a higher interest rate environment, investors just aren't paying growth-stock multiples for companies that aren't growing like one. **Market Disruption** This is fundamentally reshuffling who wins in the AI economy. The report found tech accounted for roughly 76% of S&P 500 earnings growth in 2026 through late August — an almost absurd concentration.

But a16z's forward-looking bet is that the next wave of disruption moves beyond pure software into robotics, biotech, and healthcare. If that's right, the competitive battlefield shifts from "who has the best model" to "who can get AI embedded into factories, labs, and hospital systems." That's a much harder, much slower fight than shipping a SaaS feature — it involves regulatory approval, physical infrastructure, and trust-building with industries that have historically been AI-resistant.

It also explains why we're seeing moves like Anthropic's new India data residency support through Amazon Bedrock, which Joanna flagged — regulated industries like banking and healthcare won't touch AI tools that can't meet localization law, so unlocking those verticals requires infrastructure work that has nothing to do with model quality. **Cultural & Social Impact** The sharpest line in the whole report, for us, is this: nearly 30% of companies claim quantifiable AI impact, but only 2% actually disclose a metric they track over time. That's not a rounding error — that's a trust gap at scale.

Boards, employees, and the public are being asked to believe AI is transforming productivity, while almost nobody is showing their work. This connects directly to something we flagged earlier from Microsoft's ThinkingBox benchmark, care of Joanna — agents reporting tasks as "done" while the backend database never actually changes. Multiply that dynamic up to the corporate reporting level, and you get an economy running on vibes-based ROI claims.

Workers are told AI is making their jobs more efficient, executives are told it's driving earnings, and investors are told it justifies trillion-dollar capital commitments — but the actual measurement infrastructure to verify any of it barely exists. That's a strange, slightly unsettling place for a trillion-dollar bet to be sitting. **Executive Action Plan** So what do you actually do with this if you're running a team or a budget right now?

First, if your organization claims AI impact anywhere — in a board deck, an earnings call, an internal memo — demand the longitudinal metric behind it, not the one-time anecdote. A single quarter's efficiency claim is marketing; a tracked metric over four quarters is evidence. Second, treat infrastructure spending decisions with the same skepticism software spending finally earned.

If you're evaluating vendors or cloud commitments, ask specifically where compute is landing and whether it's debt-financed, because Broadcom's $60 billion raise for Anthropic chips tells you leverage is already in the system. Third, take a16z's robotics-and-biotech thesis seriously as a planning input, not a distant trend — if material science and physical automation are genuinely the next wave, the organizations positioning now for data infrastructure in those domains, rather than just more chatbot tooling, are the ones that'll be ahead when the capital rotation actually happens. The buildout is real.

The measurement discipline isn't — yet. Close that gap in your own shop before someone asks you to prove the ROI you've been claiming.