Everyone on Tech Twitter is yelling at each other this week. Half the timeline is screaming bubble. The other half is screaming this time is different. Both halves are wrong.

The bears point at $297 billion of Q1 venture funding, 81% of it into AI, and say “1999.” The bulls point at Anthropic doubling revenue quarter-over-quarter and say “productivity miracle.” Neither is reading the right chart. The right chart is the one nobody’s drawing: where the cash actually goes when an enterprise replaces a SaaS seat with an AI agent.

This is a margin reallocation, not a bubble. The dollars aren’t disappearing — they’re moving from one P&L (Salesforce, ServiceNow, Workday) to another (Anthropic, OpenAI, the model layer). Six charts below show the transfer in progress. Then I’ll tell you which side cracks first.

Chart 1: The $297B quarter (and the four deals that made it)

Q1 2026 venture funding hit $297 billion. That’s not a quarter, that’s a decade of pre-2020 VC, compressed. AI startups captured 81%. Four deals — OpenAI’s $122B, Anthropic’s $30B, xAI’s $20B, and one mid-Q stealth round — accounted for 63% of every venture dollar deployed on the planet.

The bear reading is obvious — capital concentration this severe is what late-cycle peaks look like. The bull reading is also obvious — when the world’s three biggest software buyers (Amazon, Nvidia, SoftBank) write a $122 billion check together, they’re not betting on hype, they’re locking up supply. Both are partially right. What both miss is who’s losing.

Chart 2: The Anthropic curve

Anthropic ended 2025 at roughly $9 billion ARR. Hit $30 billion ARR in April. CFO has now guided $10.9 billion in Q2 revenue alone — which would top all of 2025 in three months and produce the company’s first operating profit (~$559M). Dario Amodei joked at a dev conference: “I’d like some more ordinary numbers.”

For comparison: this is faster than AWS at the same revenue scale. Faster than Salesforce. Faster than any enterprise software company in history. (OpenAI disputes the comparison, arguing Anthropic’s gross-revenue accounting overstates by ~$8B. Even if you take the haircut, the slope is real.)

Chart 3: The inference cost collapse

Late 2022: running a GPT-4-class model cost ~$20 per million tokens. Today: equivalent quality runs at $0.40 per million tokens, sometimes less. That is a 1,000x reduction in 36 months. The fastest cost decline in computing history. Faster than Moore’s Law, faster than memory, faster than the bandwidth-per-dollar curve in the fiber era.

The bull reading: “prices fell, demand expands.” The bear reading: “margins are toast.” Both are right. What this chart actually says is that the model layer is commoditizing faster than the application layer can capture the value. Whoever owns distribution wins. Whoever owns the model alone might not. We saw the same logic play out in Cursor’s $32-per-Pro-user margin math — the product captures the value the model layer is pricing away.

Chart 4: SaaS headcount displacement (the receipts)

Here’s the receipt the bulls usually wave around but never read carefully. Salesforce reported in its Q1 FY26 call that it reassigned 500 customer support staff after deploying Agentforce internally — a $50M annual savings. Marc Benioff was careful to call it “repositioning,” not layoffs. Tomato, tomahto.

ServiceNow disclosed in the same earnings cycle that its own internal AI agents now resolve 90% of employee IT tickets — 99% faster than human handlers. The stock dropped 13% after-hours anyway. Wall Street understood the implication before management did.

The implication: every dollar Salesforce saves through internal Agentforce is a dollar Anthropic or OpenAI could be collecting if Agentforce had been bought off the shelf instead of built. The seat-licence economics are being rewritten in real time, and the rewriter is the model layer, not the SaaS layer. The compression of pricing and headcount across enterprise SaaS is the exact pattern we traced in our 12-SaaS ARR compression model.

Chart 5: OpenAI’s revenue per employee vs SaaS comps

This one’s the kill shot. OpenAI is doing ~$25B ARR with about 4,000 employees. That’s roughly $6.25M revenue per employee. Salesforce: ~$650K. ServiceNow: ~$700K. Snowflake: ~$800K. Even peak Meta at peak scale was $1.5M.

You don’t get 9x SaaS RPE numbers in a bubble. You get them in a paradigm shift. (And you also get them when you’re burning $30B/year on training compute, which is the bear’s footnote — but the bear should explain why the customers keep paying anyway.)

Chart 6: The Anthropic valuation re-rating

End of 2024: Anthropic valued at ~$60B. Now in talks to raise at $900B. Fifteen-x in 18 months. OpenAI did its own 4x in the same window ($150B → $852B post-money on the Q1 raise). xAI tripled.

This is where the bears get loud. Fifteen-x is bubble-shaped. So is OpenAI’s revenue multiple (~34x ARR). But Cisco at 38x in 1999 wasn’t growing 80% quarter-over-quarter. Anthropic is. The multiple is high because the growth rate justifies it — until it doesn’t.

The synthesis: it’s a margin reallocation

If that’s right, two things follow. First, the AI “bubble” doesn’t pop on its own — it pops only if Anthropic and OpenAI fail to keep eating ServiceNow and Salesforce headcount. Second, the SaaS reckoning the bears are calling “the AI bubble” is actually the SaaS bubble unwinding. Different bubble. Different victim.

What I’m probably wrong about

Two things could blow up this take. One: if model commoditization (chart 3) keeps accelerating, the model layer might not capture the value either. The margin reallocation runs through the model labs, doesn’t stop at them. Two: if open-weight models from Meta, DeepSeek, or a startup nobody’s heard of catch the frontier, the entire “who collects the new margin” question gets rewritten. Either of those would cap Anthropic and OpenAI’s valuations harder than the bears want.

I think the model layer captures most of the value for at least 3 more years. After that, who knows. The whole thing might commoditize down to the cost of the GPU plus a thin distribution fee. That would actually be the best outcome for everyone except the labs. The bears would get to claim victory on multiples. The bulls would get to claim victory on productivity. Symmetry.

The predictive close

My take: by Q4 2026, the first major SaaS incumbent — Salesforce or ServiceNow, not both — announces a meaningful headcount cut and frames it as “AI productivity gains.” That’s the moment the market finally re-prices SaaS, and the model-layer multiples look reasonable rather than insane. Bears get a victory lap on whichever incumbent cracks. Bulls get to keep their AI bags. Everyone else gets cheaper inference and worse customer support. Story of our time.

This isn’t 1999. It’s 2007, and the model labs are the iPhone. The seat-licence economy is the Nokia. Watch which incumbents announce “AI strategy” press releases this fall — those are the Nokias.

Sources: Crunchbase Q1 2026 venture report, VentureBeat on Anthropic ARR, CNBC on Q2 guidance, ServiceNow Q1 2026 earnings coverage, Salesforce Q1 FY26 earnings, TokenMix API pricing history. Revenue-per-employee figures are illustrative composites from public reporting. Trial and event photographs are reproduced for editorial commentary under fair use.