Enterprise software just went through its sharpest repricing since the dot-com era. The headline is that AI is eating SaaS. The reality is more specific — and from where we sit, writing the first checks into YC companies before Demo Day, more interesting. Here is what the public market actually repriced, what the AI buildout looks like on both the supply and demand side, and what the earliest-stage pipeline tells us about where this goes next.
Three takeaways
- The market repriced software; it did not reprice demand. The drawdown took roughly $2 trillion off enterprise software at the trough. What actually broke is per-seat pricing, not SaaS as a delivery model — total enterprise software spend is still on pace to rise ~15% to $1.4 trillion in 2026. The money is moving, not leaving.
- We sit at the entry point of that migration. The YC batches we deploy into are the supply side of the repricing: agent infrastructure, outcome-priced software, and the tooling that lets small companies replace seat licenses entirely.
- The AI buildout is a bubble on several metrics — and our book is on the right side of it. Roughly $700 billion of hyperscaler capex this year has run ahead of proven returns. But demand is the stronger story, and our portfolio is a capex-light, demand-side bet whose input costs fall as the buildout overshoots.
What the Repricing Actually Hit
A February panic wiped over $1 trillion off software stocks in a single week; agentic product launches triggered a $285 billion single-day wipeout; and software forward multiples fell below the S&P 500 for the first time on record. By mid-2026 the index-level panic is over — the software basket is back green for the year — but that is a violent bounce off the spring bottom, not a return to old highs, and it is wildly uneven.
The dispersion is the whole story. The AI picks-and-shovels layer got paid — DigitalOcean, Datadog, CrowdStrike, and Snowflake all ripped off the lows as earnings showed the best incumbents can monetize AI rather than be eaten by it. The seat-based application names caught the bounce but remain deeply repriced:
| Company | 2026 move | What the tape is debating |
|---|---|---|
| Salesforce | ~30–38% down YTD | Agentforce growth is real but small against the seat base; the market is pricing seat compression before it shows up in revenue. |
| ServiceNow | ~36% down YTD | 20%+ subscription growth, 97% renewals, AI target raised from $1B to $1.5B in one quarter — and half of net-new business is already non-seat pricing. |
| SAP | ~33% down YTD | Owns the ERP data agents must connect to; shipping 50+ domain agents; AI infra costs hit margins before AI revenue ramps. |
| Figma | ~52% down (H1) | 46% revenue growth and 139% net dollar retention — and the stock still halved. A pure narrative trade on seat erosion. |
| DocuSign | 40%+ down (1yr) | $1B+ free cash flow, no debt, buybacks; pivoting from e-signature point product to an agreement platform. Classic thin-moat repricing. |
| Zoom | Derated with the sector | A ~4% grower repositioning as a workflow system of action; mature per-seat communication tools are seen as the most exposed shelf. |
| Indian IT (Infosys et al.) | Sector at 3-year lows | AI attacks the labor-arbitrage services model directly — exactly where YC is aiming. |
What is being repriced is the unit of pricing, not the category. Old SaaS charged a fixed price per human seat. AI-era contracts charge for work done — per token, per task, or per resolved outcome — because an agent does the work of many seats without holding a license. This is already in market: Intercom charges $0.99 per AI-resolved ticket, Zendesk $1.50–$2.00 per automated resolution, and Salesforce prices agents on completed actions. Hybrid models (platform fee plus usage) are now the dominant structure at 61% of SaaS companies, up from 49% in 2024.
At the top end, agent-as-FTE pricing runs $800–$2,000+ per month per agent, anchoring against a ~$60,000 salary instead of a $20 seat license. That anchor shift is why total software spend keeps rising even as seat counts compress. Meanwhile at the small end, SMBs are pulling spend out of SaaS entirely — replacing Salesforce and HubSpot contracts with apps they build on AI tooling and cutting software costs 40–80%. One 20-person shop replaced a $40,000 Salesforce contract with a self-built CRM costing $1,200 a year. Gartner puts the spend exposed to this agentic arbitrage at $234 billion by 2030.
The bottom line: SaaS as a delivery model is not dying — per-seat pricing is. The durable moat is proprietary data, system-of-record ownership, and the compliance and reliability bar that makes DIY replacement infeasible for mission-critical work. That is why ServiceNow and SAP screen as survivors, why thin-moat seat-priced point products stay exposed, and why IT services face the most direct assault.
A Bubble on the Supply Side — With Demand Intact
Is the AI buildout a bubble? On several metrics it already is — and that matters less for our strategy than most assume. The supply side is historic: the five largest US hyperscalers have guided to roughly $700 billion of combined 2026 capex, up ~77% on 2025. Goldman's baseline implies ~$765 billion of AI capex this year growing to ~$1.6 trillion by 2031.
The bubble evidence is real. The AI Big 10 now make up ~41% of the S&P 500 — comparable to tech and telecom at the dot-com peak. Moody's counts roughly $662 billion of signed-but-not-commenced data-center leases sitting off balance sheet. Depreciation schedules of five to six years sit on silicon with a two-to-three-year economic life. Valuations imply near-perfect execution, and a rationalization phase will cull weak use cases when contracts renew.
The demand side gets less attention, and it is the stronger story. Anthropic reached a ~$47 billion annualized run rate, up from roughly $10 billion for all of 2025, with more than 1,000 enterprise customers spending over $1 million a year. Per-user economics rhyme: the $20 plan anchors the market, but $200–$300 premium tiers are the fastest-growing segment, because the frontier model saves hours and hours are worth more than tokens.
Agents explain why bills rise as unit prices fall. A multi-step agentic workflow burns 10–50× the tokens of a simple chat, so total spend climbs with capability. When one large employer's engineers burned an entire annual AI-coding budget in four months, the response was to cap spend per engineer — rationing a resource they want more of, not walking away from it. Token prices are flattening while compute costs keep falling 60–70% a year; stable prices on falling costs is the definition of pricing power. Underneath it is labor substitution: when software replaces legal review, coding hours, and research analysts, the reference price stops being other software and becomes a salary, and elasticity collapses.
Where we land: a correction in AI equities and infrastructure spending is entirely possible without the demand thesis being wrong. The market still prices AI like software. Users have started pricing it like staff. That gap can keep repricing long after the froth clears.
What the YC Pipeline Shows
Everything above is visible on a Bloomberg terminal. What follows is not. We write pre-Demo Day checks into every YC batch, which gives us a look at the AI-native supply side roughly 18 months before it shows up in anyone's competitive analysis. We ran our own founder-level analysis of the recent batches and cross-checked it against the public YC company directory.
Composition: B2B cedes, hard tech rises. B2B software declined from 64% of the Winter 2026 batch to 61% of Spring 2026 to 55% of the Summer 2026 companies announced so far, while Industrials climbed to roughly 22% of Summer 2026 — its highest share in recent memory. Within B2B, the concentration is in infrastructure and engineering tooling: the picks-and-shovels layer of the agent economy.
| Segment | W26 | P26 | S26* |
|---|---|---|---|
| B2B software | 64% | 61% | 55% |
| Industrials (deep tech) | 14% | 12% | ~22% |
| Fintech | 9% | 10% | 6% |
| Healthcare | 8% | 9% | 7% |
| Consumer | 4% | 6% | 2% |
| Real estate & construction | 2% | 2% | 5% |
Founder quality: operators in, consultants out. Founder pedigree is deep — among founders with a known school, roughly half attended a top-tier university, led by Stanford, Berkeley, and MIT, with prior employment concentrated at Amazon and Google. Two shifts stand out. Consulting pedigree is thinning fast (McKinsey founder counts fell 11 → 8 → 2 across the three batches), and a hard-tech employer cluster recurs every batch: SpaceX, Tesla, Palantir, NASA JPL, and Scale AI alumni are the talent base behind the Industrials climb.
The cohorts are also getting younger and more first-time. Average founder age fell from 26.3 to 24.1 batch-over-batch; prior-YC-alumni founders dropped below 2%; and two-person teams remain the modal structure despite noise about an AI solo-founder surge.
| Metric | W26 | P26 |
|---|---|---|
| Companies | 198 | 196 |
| Founders | 436 | 434 |
| Average founder age | 26.3 | 24.1 |
| Average team size | 2.2 | 2.2 |
| Solo-founder companies | 8% | 11% |
| Prior YC-alumni founders | 3.7% | 1.6% |
| US-based companies | 84% | 96% |
Read plainly: YC is betting overwhelmingly on first-time founders entering on accessible AI tooling. Talent density stays high while operating experience softens. For a concentrated picker, that is a problem. For a strategy underwriting the batch-level hit rate, it is the environment working as designed.
How We're Positioned
- Deploying into the current batch. Recent checks include TesterArmy, Arga Labs, Chronicle Labs, and Standout — all in the agent-infrastructure and reliability theme that defines the batch and that the public market is now paying for.
- Sourcing weighted to infrastructure and outcomes. Toward B2B infrastructure and engineering tooling, and toward companies that can dollarize ROI rather than sell seats. The renewal-driven rationalization ahead will cull use cases that cannot show the math.
- Watching the Industrials cluster. At ~22% of the latest batch it is too large to ignore, and the recurring SpaceX / Palantir / JPL talent pipeline suggests durability. Deep tech carries longer feedback loops, so we size accordingly.
- A demand-side bet, not an infrastructure bet. Our seed book is capex-light software consuming compute whose cost falls 60–70% a year — effectively subsidized by the hyperscalers' buildout. A capex correction lowers our companies' input costs while it punishes the infrastructure-levered names.
What Could Go Wrong
- Softer exits. Compressed incumbent stock prices mean cheaper acquirer currency and a weaker M&A bid for seed-stage software — and our most likely exit path runs through the very companies whose stocks just halved.
- Seed valuations correcting with the narrative. The spring rebound showed how fast sentiment mean-reverts; if the disruption story cools, some froth in AI-native seed valuations comes out with it.
- The rationalization cull. When enterprise AI contracts renew, weak use cases get cut. Companies without provable, dollarized ROI are exposed — which shapes our sourcing filter directly.
- Execution risk from younger, first-time founders. Less operating experience per company, batch over batch. Diversified batch coverage is the mitigation, not a cure.
- A hard capex unwind. Falling compute costs help us, but a disorderly infrastructure correction would freeze the venture funding environment — including our companies' follow-on rounds — regardless of demand-side fundamentals.
This is market commentary, not investment advice. Public-market figures are approximate and several names have partially recovered; batch figures reflect our own analysis of the YC cohorts and are subject to change.
