Last week, Palo Alto Networks CEO Nikesh Arora said something that made a lot of tech executives sit up straight. AI pricing, he told CNBC, needs to fall by as much as 90% before enterprises will seriously adopt it at scale. Not 10%. Not 30%. Ninety percent.
That’s a staggering number. And it lands right in the middle of what’s becoming the most important — and uncomfortable — conversation in tech right now: does AI actually pay for itself?
Let me unpack why a 90% price cut isn’t just wishful thinking from one CEO, and why it might actually be the most reasonable take I’ve heard all month.
Follow the Money — $1.5 Trillion in Chips, $3 Trillion in Questions
A few days before Arora’s interview landed, Sequoia partner David Cahn updated his famous “Are we building too much AI infrastructure?” math. You might remember his original 2023 note — back when Nvidia’s GPU revenue was $50 billion, he calculated the industry would need $200 billion in AI revenue to justify the spending.
Fast forward to today. That number has ballooned to $1.5 trillion in infrastructure spend for 2026. And the required revenue to justify it? Cahn’s estimate is now $3 trillion.
Let that sink in for a second. Three. Trillion. Dollars.
To put that in perspective, the entire global software industry — every SaaS product, every enterprise license, every operating system — generates about $1 trillion annually. AI alone, by Cahn’s math, needs to be three times bigger than the entire software industry just to break even on the hardware we’ve already ordered.
Anthropic is reportedly doing around $60 billion in ARR. OpenAI earned roughly $13 billion in 2025, though they claimed $20 billion ARR by November. Those are huge numbers for startups. They’re a rounding error compared to $3 trillion. And as I noted recently about Microsoft quietly replacing OpenAI in its own apps, even the biggest enterprise customers are rethinking their AI spending.
The Token Price Problem Nobody Wants to Talk About
Here’s where Arora’s comment starts to make more sense. The fundamental unit of AI — the token — is both getting cheaper and being used more efficiently. OpenAI’s Sam Altman recently noted that GPT-5.6 is 54% more token-efficient on coding tasks than its predecessor. That’s great for my monthly API bill. But it’s terrible for the “sell more tokens” business model.
Arora’s argument isn’t just that tokens are too expensive. It’s that at current prices, the economics don’t work for the kinds of enterprise use cases that would actually drive mass adoption. A company might spend $100,000 a year on AI tools today. For AI to truly transform how that company operates — embedded in every workflow, every customer interaction, every decision — that number needs to work at enterprise scale. And at current prices, it doesn’t.
This isn’t academic for me. As someone who manages IT budgets in a government setting, I run into this tension every quarter. The tools are genuinely impressive. The cost projections for scaling them across an organization? Less impressive. This echoes something I wrote about how better models don’t always mean better tools — the gap between what AI can do and what it can sustainably deliver at scale is wider than most vendor demos suggest.
The Hyperscaler Gambit
Torsten Slok, chief economist at Apollo, recently flagged a risk that should worry anyone who’s watching the market: the hyperscalers — Google, Meta, Microsoft, Amazon — are all projecting massive free-cash-flow accelerations in 2028. That’s their bet: spend billions now (on chips, data centers, energy), reap trillions later.
Slok’s warning is simple but chilling. What if that inflection point doesn’t arrive on schedule? Or at all? A slower AI payoff wouldn’t just be a sector problem, he writes. It “would risk tipping the economy into recession and the S&P 500 into a correction.”
That’s not a small footnote. That’s a systemic risk hiding in plain sight, masked by all the excitement around what AI could do versus what it’s actually generating in revenue today. The AI hype cycle is showing real signs of strain, and this is the economic pressure behind it.
To be fair, there are genuine revenue streams emerging. Anthropic and OpenAI are both profitable now. Enterprise AI adoption is accelerating. But the numbers need to get three orders of magnitude bigger to close the gap Cahn identified. That’s not a growth curve — that’s a hockey stick that needs to turn into a vertical line.
What This Means for Developers and IT Leaders
So where does this leave those of us who are actually using AI day to day?
On one hand, the tools have genuinely changed how I work. I use AI coding assistants daily — writing this very post wouldn’t be the same without them. The productivity gains are real. I’ve seen junior team members ship code that would’ve taken them weeks a few years ago, in days. The value proposition for individual developers and small teams is not in question.
On the other hand, I’ve also seen the bills. Running a moderately active AI-augmented workflow — agents, embeddings, code generation, the whole stack — adds up fast. For a government ICT office in the Philippines, where budgets are scrutinized down to the peso, the ROI conversation is very, very real. I looked at how AI is reshaping the job market for developers, and the same economic pressures are hitting individual careers too.
Arora’s 90% price cut target sounds extreme until you run the numbers yourself. For AI to truly become infrastructure — like electricity or internet connectivity — it needs to be cheap enough that you don’t think about the cost. You just use it. We’re not there yet.
The Open Source Wildcard
One force that’s already driving prices down is the open source AI ecosystem. Ollama just raised $65 million and hit 9 million users. Models like Grok 4.5 are pricing at $2 per million input tokens versus Anthropic’s $5 or OpenAI’s higher tiers. I broke down the Grok 4.5 pricing and benchmarks — it’s genuinely competitive, and that kind of competition is what ultimately drives prices down.
The great deflation in AI pricing is already happening. It just hasn’t reached the numbers Arora is talking about yet. The question is whether the hyperscalers can sustain their spending long enough for the market to catch up — or whether we hit an “AI winter” first as the expectations reset.
I’m cautiously optimistic that the deflation will accelerate. More competition, better architectures, and the relentless march of hardware improvements (AWS Graviton5 is 30% faster than Graviton4, and Samsung is reportedly working on a dedicated AI accelerator chip for PCs) all point toward cheaper inference over time. But “over time” and “by next quarter” are very different timelines.
My Bottom Line
The $3 trillion question isn’t really about whether AI is valuable. It’s about whether the current business model around AI — centralized, capital-intensive, token-metered — can generate enough revenue to justify the bet that Silicon Valley has placed on it.
As someone who builds software and manages technology for a living, I think the answer is more nuanced than either the hype merchants or the doomers want to admit. The technology itself is transformative. I’m not sure the current pricing model is sustainable, and I’m even less sure that the hyperscalers’ revenue projections will hit on schedule. But the gap between where we are and where we need to be is closing — just not at the pace the infrastructure spend would suggest.
If you’re an IT leader evaluating AI investments right now, my advice is simple: invest in the capabilities, not the hype. Use AI where it genuinely moves the needle. Track your actual costs. And keep an eye on the pricing trends — because in 12 months, the economics of AI might look very different than they do today.
Whether that’s because prices finally dropped to where enterprises can adopt freely, or because the correction arrived first… well, that’s the $3 trillion question.