a16z Podcast

Andy Jassy on why $220B still leaves AI demand on the table

Andy Jassy· President and CEO of Amazon at Amazon
·~61 min·English·a16z
AI CompanyInferenceBusiness StrategyAI Infrastructure
TL;DR

Amazon's CEO argues AI demand so far outruns supply that even $220 billion in 2026 capex leaves demand on the table — and that meeting it means rebuilding the whole stack, from custom chips that cut the cost of inference to org charts, security, and products reimagined from scratch.

01The Demand Gap

Leaving Demand on the Table

Amazon will spend $220 billion on capex in 2026 and still can't meet AI demand — 2026 is sold out, 2027 largely so, and chunks of 2028 are already gone.

Even spending $220 billion of capex this year, we're leaving a meaningful amount of demand on the table.

— Andy Jassy, a16z Podcast
Key Insight
The sold-out years are a demand signal a competitor can't fake: Amazon is pre-selling capacity it hasn't finished building, which points to a physical constraint — the power, cooling, and chips Jassy says the labs will pay almost anything for — rather than a shortage of money. That reframes Wall Street's dot-com fear: the risk he describes isn't empty data centers, it's not having enough of them.
02The Demand Barbell

The Demand Barbell

AI demand is a barbell: compute-hungry labs at one end, cost-cutting enterprises at the other, and the largest prize — ordinary enterprise apps not yet running inference — still sitting untouched in the middle.

In the middle of the barbell are the current production workloads of enterprises.

— Andy Jassy, a16z Podcast
Key Insight
The barbell points to where Amazon expects its largest future segment: not the labs, but the middle — the ordinary enterprise apps Jassy says still aren't using inference in any pervasive way. On his own logic that middle only converts once inference gets cheap, which is what makes the chip strategy in a later section the hinge of the whole thesis.
03The Capex Bet

Two Bets, One Capex Number

Amazon splits the capex: the 30-year data-center shell is placed two years ahead of demand, but the fast-depreciating servers are bought only months out, once customer demand is actually visible.

So, you have very good visibility into customer demand before you pull that trigger.

— Andy Jassy, a16z Podcast
Key Insight
The asymmetry is the whole risk model. The speculative part — the building shell — is cheap per year across a 30-year life and placed early; the expensive, fast-aging part — servers — is bought only when demand is already on the books. So "are you spending too much?" is really two questions, and only the smaller, reversible half is a bet on the future.
04The Infrastructure Case

Data Centers Are the New Highways

Jassy argues data centers are the biggest infrastructure buildout since the 1950s highways, and that the loudest objections — water, pollution — don't survive the numbers.

So, we have to build data centers.

— Andy Jassy, a16z Podcast
Key Insight
Framing data centers as infrastructure rather than industry casts the buildout as a national necessity, the same framing that once justified public highway spending. The counter-numbers he offers — golf courses using 200 times the water, backup generators running about ten hours a year, billions in tax to host counties — are each answers to a specific objection, which tells you the fight he expects to win is the one over facts rather than values.
05The Cost of Inference

Inference Is Still an Order of Magnitude Too Expensive

The real growth lever isn't the lab deals — it's driving the cost of inference down an order of magnitude, which is why Amazon builds its own chips, Graviton and Trainium.

But if you can take the cost of inference down, that bar is much lower and all applications are going to be better and smarter and more capable with AI accelerators and inference.

— Andy Jassy, a16z Podcast
Key Insight
By naming the cost of inference — not the lab deals — as the thing that unlocks pervasive enterprise use, Jassy ties Amazon's largest opportunity to its own chips. Building Graviton and Trainium for 30-40% better price-performance is how Amazon pursues that cost curve itself rather than depending entirely on a single dominant supplier, which is the connection he draws between the silicon and the middle of the barbell.
06Org Design

Six Engineers, Seventy-Six Days

AI compresses the org chart: the interview cites a Bedrock inference engine rebuilt by six engineers in 76 days — once a 40-person year — and Jassy expects fewer humans doing today's knowledge work.

we're going to have fewer humans doing what they've done today

— Andy Jassy, a16z Podcast
Key Insight
"Fewer humans" here is about knowledge work, not the warehouse — Jassy pairs it with smaller, cross-functional teams and models that already get a small group 80% of the way there. The implication is that the harder thing to change may be organizational rather than technical: the specialization incumbents built up over decades is exactly what the small-team model asks them to unwind.
07The Real Risk

The Risk Is Cybersecurity, Not Sci-Fi

The risk that matters now isn't a rogue superintelligence — it's cybersecurity, where AI finds vulnerabilities faster than humans ever did and patches must ship in minutes.

I don't think that there is enough discussion about the more immediate term risks, which are really around cybersecurity.

— Andy Jassy, a16z Podcast
Key Insight
By calling the sci-fi scenarios "possible" but cybersecurity "immediate," Jassy steers the safety debate toward the threat AWS already sells a product for. His underlying point still holds on its own terms: if vulnerabilities surface in minutes, the response has to come in minutes too — which he says is impossible unless the application was rebuilt so agents can update it, forcing the rewrite in the next section.
08Reinvention

Reinvent the Brain, Don't Repaint It

Incumbents that treat AI as an incremental upgrade will lose; the winners rebuild each experience from scratch, the way Amazon rewired Alexa's brain.

you have to find a way to get teams to reimagine from scratch what those applications look like if you were building it today with all the new technology available like AI

— Andy Jassy, a16z Podcast
Key Insight
The hard part isn't knowing you must rebuild — it's that the people best at the old system are the ones who must tear it down. Jassy's real claim is organizational, not technical: the incumbents most likely to lose are the ones whose pride in what they built blocks the restart, which is why he frames it as change management, not engineering.