This week’s insight
Better models created the gross margin problem. Usage-based pricing and open models will solve it.
Josh Trup, Principal
One of the less obvious consequences of better frontier models this year has been what they did to application-layer economics. Better models made products more useful, which drove more usage, but many AI applications were still charging like traditional SaaS on a fixed seat while paying for inference on every token. Agentic workloads make that mismatch much bigger: OpenRouter estimates an agentic request uses roughly 15x more tokens than a human-driven one. The slightly counterintuitive result is that a product getting better and being used more can actually make its gross margins worse.
We’re now seeing the response. Rather than sending every task to the most capable frontier model, applications are increasingly routing routine work to cheaper open-weight models and reserving frontier models for the tasks where the quality premium actually matters. Open-weight models went from 7% to 56% of token volume on Vercel’s AI Gateway between December and August, while accounting for only 14% of spend. On OpenRouter, DeepSeek doubled its token share from 9% to 18% in the first half of the year as V4 became good enough for agentic workloads. The economics are hard to ignore: V4 Flash was priced at $0.09/$0.18 per million input/output tokens versus $5/$30 for GPT-5.5.
Harvey gives a glimpse of where this could go. It is now post-training open-weight models specifically for legal work and has shown internal benchmarks where specialised models perform competitively with frontier models at dramatically lower cost, including one workflow at roughly one-tenth the cost per cell and another with 90% lower cost per query. At the same time, OpenRouter found that temporarily discounting OpenAI’s models drove token usage up 5.6x and 13.8x, with nearly a third of users continuing to use them after the discount ended. If intelligence keeps getting cheaper and applications can hold their own pricing, the application layer could get the best of both worlds: substantially more usage and structurally better gross margins. The important question may become less which frontier model wins, and more which applications control the routing layer and keep the savings.
The charts are updated 👇
Israeli High-Tech Funding: Cumulative YTD Tracker
SaaS Public Comps
Top 10 by EV / NTM Revenue
Ranked by EV / NTM revenue multipleas of
SaaS EV / NTM Revenue vs. NTM Revenue Growth
2025-26 SaaS IPOs added:
SaaS EV / NTM Revenue Multiples Over Time
AI Infrastructure
AI Infra EV / NTM Revenue vs. Rule of 40
Every cohort constituent is plotted. Bubble size scales with market cap. Per-name detail for each cohort is in the tables below.
AI Infra EV / NTM Revenue Multiples Over Time
This week in AI infrastructure
Large-Cap Security
Large-Cap EV / NTM Rev vs. Growth
Security Comps
Sorted by enterprise valueas of
Macro Indicators
Macro-Economic Indicators
Sources and method. Public market data from stockanalysis.com (S&P Global Market Intelligence and Nasdaq Data Link), as of October 2, 2026. The two EV / NTM Revenue Multiples Over Time charts are built from PitchBook consensus data. TSMC's EV / NTM Revenue is taken from PitchBook rather than stockanalysis, so the figure in the Semiconductors table matches the multiples-over-time chart; its remaining columns are stockanalysis-sourced like every other name. EV / NTM Rev is enterprise value divided by consensus next-fiscal-year revenue. Growth-adjusted multiple is EV / NTM Rev divided by NTM growth. Rule of 40 is NTM revenue growth plus LTM free-cash-flow margin. P/E shows "n/m" where earnings are negative. SG&A percent of revenue is shown because the data source does not break out a separate sales and marketing line. Year-to-date trading is an ETF-style, market-cap-weighted cumulative return from January 2026, using a fixed basket at current share counts. P/E is the market-cap-weighted trailing P/E of profitable constituents, shown on a log scale. Capex is total reported capital expenditure, as companies do not disclose an AI-only figure. LTM is the sum of the four most recently reported quarters of cash capital expenditure, from each company's cash-flow statement (source: stockanalysis.com), and it updates every quarter rather than once a year. Fiscal calendars differ, so the LTM windows are not identical: Microsoft, Amazon, Alphabet and Meta run through June 2026, Oracle through May 2026. The 2026 column is company guidance or consensus (Amazon $200B, Alphabet $185B, Meta $135B). Microsoft and Oracle have already closed their 2026 fiscal years, in June and May respectively, so both show actuals rather than estimates: Microsoft $115.9B and Oracle $55.7B, which is why each matches its LTM bar. Microsoft's widely quoted FY26 figure of $190B, later revised to roughly $175B, counts capital expenditure plus finance leases; this chart uses cash capital expenditure from the cash-flow statement throughout, so the two are not comparable. Guidance for the calendar-year companies is not always stated on that same basis, so their 2026 bar and LTM bar are not strictly like for like. SpaceX issued no formal 2026 capex guidance; its bar is H1 actuals of $28.5B plus management's remark on the Q2 call that quarterly capex stays near $18.4B, so it is a derived figure rather than a company forecast. Bubbles on the EV / NTM Rev and Rule of 40 charts are sized by market cap, scaled so that area rather than width tracks it. Sizes are capped above roughly $680B and below roughly $19B, so the very largest names are drawn alike and so are the very smallest; hover any bubble for the figure. For the security chart, NTM Revenue Growth is used as a proxy for ARR growth. The 2025-26 SaaS IPOs (SailPoint, Figma, Navan, Netskope, and MNTN) are added to the scatter; none currently price above the EV / NTM Rev top ten. Macro indicators: the 10-year Treasury yield, the effective Federal Funds rate and CPI year-on-year, monthly from January 2012 (macro tracker workbook, PitchBook-sourced). This is not investment advice. Verify before acting.