The market map

The idea in one line: In the six weeks after Jev launched, the generic decision model became a crowded, nearly free commodity, and the money and the exits went to whoever owns the data, the verifier or the distribution.

A market map tells you which layer of a business you can win from a laptop and which layers are already owned. That choice is a 10x choice, not a 10% one. At October 2026’s price floor, a 2-billion-parameter decision model can keep about 87% gross margin while a 27-billion-parameter one loses money on every call (an estimate, worked through below). Tabular model makers were bought by the companies that own the data platforms. Biology labs raise a billion dollars because they run wet labs.

This lesson draws the map as of October 2026; wedges, moats and proof turns it into a plan.

A Request for Startups (RFS) is Y Combinator’s public list of ideas it wants founders to apply with. YC says the list is only a fraction of what it funds.

The quote that frames this course is the RFS entry “Vertical Foundation Models”, signed by Diana Hu, on YC’s Winter 2027 list (YC Requests for Startups, read 2026-10-28). It opens: “I think Jev hit a nerve because it gives developers a typed decision and a probability instead of generated text.” It ends with the bar: “You’ll need rights to the data and a clear win in accuracy and compute over existing specialist models.”

• Who. Diana Hu became YC Managing Partner on 2026-06-11. She was CTO of Escher Reality (YC S17, sold to Niantic) and studied computer vision and ML at CMU (YC blog, 2026-06-11). She argued the same thesis in March 2024, profiling 25 YC companies that trained or fine-tuned their own models inside one three-month batch on $500k plus over $1M in cloud credits (YC blog, 2024-03-28).
• When. The page shows no date. One source says Wayback Machine captures show the Winter 2027 list first on 2026-10-27 (FourWeekMBA, 2026-10-28); only that one source says so. The page listed an application deadline of November 2.
• The siblings. Eleven entries sit on the same list. Three matter here:
  • Own Your Intelligence (Jon Xu): companies will want to own specialized models trained on their own data, and rent frontier models only for the hardest problems. This is the founder-side twin of Hu’s entry.
  • AI for the Air-Gapped Economy (Brad Flora): “The most valuable data in the world isn’t in the cloud.” It asks for on-prem models for defence, banks, hospitals and nuclear sites.
  • LLMs That Write Well (Jared Friedman): as models improved at code and maths they got worse at writing. That reads as an admission that verifiable rewards shape models in ways nobody intended.
• The read. FourWeekMBA counts seven of the eleven as about hardware, financing, government and work: YC thinks the bottleneck has moved past the models.

Both examples the entry names are real, recent and benchmarked by their own makers. Jev launched on 2026-09-15 (The Register, 2026-09-16). TimesFM-3 weights went public on 2026-08-31 under a non-commercial licence (Pillitteri, 2026-09-01), so “cheap enough to run on everything” does not yet apply to it in production unless you pay Google.

Start with the plain idea. A decision model reads messy input like a language model but answers like a classifier: a typed choice, score or yes/no, with a probability for each allowed answer, from one pass and without writing text. Anatomy of Jev covers the contract and generating vs deciding covers why skipping text generation is so much cheaper.

Now watch what happened once one company proved people wanted it.

Sources: VKTR, 2026-09-16; InfoWorld, 2026-10-13; UsagePricing, 2026-10-15; ToolNavs, 2026-10-21. By 2026-10-28 the Decision Index data listed 243 entries (Decision Index).

Copying was literal. Cloudflare made Clef answer Jev’s API so developers can switch without rewriting, and reused the name of TypeSafe’s training method, “Reinforcement Learning for Calibrated Decisions” (RLCD), for its own pipeline (remio, 2026-10). The three answer types (yes/no, choice, ordered score) became an industry convention within a month (lilting.ch, 2026-10-06).

Notice who is not on the list. No dedicated Gemini decision endpoint was found as of 2026-10-28. Anthropic’s structured outputs (generally available 2026-02-04) guarantee the JSON shape but return no per-option probabilities.

Decision APIs charge per input token (a token is a word piece, roughly three-quarters of a word) and give the output away, because the output is a handful of numbers. So the bill is input length times number of calls.

Worked example. Microsoft’s Decision-1 costs $0.042 per million input tokens. An 86-token request costs 86 × 0.042 / 1,000,000 = $0.0000036, so $1 buys over 275,000 decisions (InfoWorld, 2026-10-13). Compare a general LLM on the same job: 500 input tokens and 30 output tokens on GPT-5.6 Terra at $2/M in and $12/M out is $0.001 + $0.00036 = $0.00136. Jev on 500 tokens is $0.000021. That is about 65x. Arize measured a wider gap on real judgments: $14.30 per 1,000 for Opus 5 against $0.045 for Jev, about 318x (Arize, 2026-10-07).

The list prices, as of October 2026:

Decision API list price, $ per 1M input tokens, output free
(as of October 2026; one # = $0.01)

Clef 27B, Cloudflare*        0.240  ########################
GLiDE, Fastino               0.150  ###############
OpenAI Decisions             0.100  ##########
Jev, Decision-1, d1-ultra    0.042  ####
GLiNER2.5-Decide, Pioneer    0.030  ###
Perplexity Decider (cut)     0.020  ##
Perplexity Decider Flash     0.010  #

* sources disagree on Clef-flash: $0.09 or $0.24

Perplexity cut its price eleven days after launch, then cut again. Three vendors sit at exactly $0.042. That convergence is what a commodity looks like.

The floor also sits close to the cheapest general LLMs. One estimate, at 500 input and 30 output tokens per decision, puts Jev at $21 per million decisions against $35-65 for the small frontier tier (Claude Haiku 5.5, GPT-6-luna, Gemini 2.5 Flash-Lite) and $1,300 for Sonnet-class models. Batched or cached, the small tier comes within about 1-3x of Jev. “Orders of magnitude cheaper” is true against flagship models, not against the cheapest ones, so a decision vendor must also sell calibration, latency and privacy.

Adoption did not turn into revenue. Vercel said nearly 13% of paid AI Gateway teams used Jev within 24 hours, about twice the GPT-5.6 family’s share (InfoQ, 2026-10-01). Two weeks in, AI Weekly, citing TechCrunch and Vercel, put it at 35% of paid teams but only 0.27% of gateway spend (AI Weekly, 2026-10-17); that figure comes from a single source. Broad reach, tiny revenue per call: that gap is the business problem every decision vendor now has.

Vendors grade their own homework. The best outside head-to-head so far is Arize’s (Arize, 2026-10-07). Calibration error here is the average gap between how confident a model says it is and how often it is right; lower is better (see calibration). p50 latency is the median response time.

The top three are within about one point on accuracy. What separates them is calibration, latency and price. Arize also flipped the question’s wording. Jev, Clef and OpenAI held their ROC AUC (how well a model ranks right answers above wrong ones; 0.5 is a coin flip, 1.0 perfect) at 0.89-0.93, while Liquid d1 collapsed to 0.17, worse than a coin flip. Calibration on one wording does not carry over to another.

Other outside evidence points the same way:

• Vals found Jev within noise of frontier models on 400 claim checks against 10-K filings (97.5%) but last of 12 on a LegalBench slice (73.0%) (Vals AI, 2026-10-06).
• In Every’s narrow test, Jev caught 6 of 7 planted defects and Fable 5.1 caught 7 of 7 (SmartScope, 2026-09-19).
• Bryo AI found Gemini slightly more accurate on email classification but 10-20x more expensive (InfoQ, 2026-10-01).
• Leaderboards disagree with each other. JevBench 1.5 puts two small-team entries, Cygnet (73.70) and Winnow-12B (73.23), in a statistical tie above Jev 1.13 (72.13) (BenchLM). On the Decision Index data of 2026-10-28, 47 entries outscore Jev 1.14. BenchLM warns the editions are not directly comparable.

The founder’s reading: the hosted leaders are level on accuracy, open clones are close behind, and no one can win on “our model is smarter” in a generic market.

A tabular foundation model is trained once on millions of small, mostly synthetic tables, then predicts on your table from the rows you hand it, with no training run (TabPFN, TimesFM and friends). A relational one does the same across a database’s linked tables.

Both stories ended the same way in 2026:

• SAP agreed to buy Prior Labs (maker of TabPFN) on 2026-05-04. SAP did not disclose the price, but pledged over €1B across four years. Prior Labs had raised a €9M pre-seed in February 2025, so the path from pre-seed to acquisition took 15 months (SAP News, 2026-05-04). Headlines calling it a “€1B deal” mix up the price with the investment pledge (tech.eu, 2026-07-17).
• Nvidia bought Kumo (KumoRFM) in mid-2026, with the cofounders moving over in May (Fortune, 2026-06-03). The reported price of about $400M rests on one anonymously sourced report by The Information (PYMNTS, 2026-06-04). Kumo had raised $37M in 2022.

The pattern: the buyer owns the platform where the data lives. SAP owns enterprise resource planning data. Nvidia sells the compute and wants the enterprise workload. For a founder, that names your likely acquirer before you write a line of code.

Money follows the cost of the verifier. A verifier is whatever tells you an answer was right. In biology and materials it is a wet lab, which is expensive. In forecasting it is tomorrow’s data, which is free.

Caveats matter here. The TypeSafe Series A comes from TypeSafe’s own announcement (TypeSafe, 2026-10-09). Chai’s valuation comes from one secondary page with internal inconsistencies. Lila’s and Periodic’s 2026 rounds were reported as in talks, not closed. Silurian’s $6M round is single-source.

So do not start a general protein, materials or weather model: those markets are funded at $100M-$2B or given away by Google, Datadog, Amazon and Arc.

Open weights means you can download the trained model. The licence says what you may do with it. Two camps formed.

• Open for research, paid for production. TabPFN-2.5 through 3.5 and TimesFM-3 ship non-commercial weights. Prior Labs’ terms bar commercial use including derivatives and outputs used for business decisions; production needs the API or a commercial licence (Prior Labs docs, read 2026-10-28). TimesFM-3’s commercial route was BigQuery, announced as “coming weeks” in September; its status at 2026-10-28 is unverified. The pattern keeps the academic download flywheel and stops cloud hosts reselling the model. It worked well enough that SAP bought the company.
• Apache 2.0. Datadog Toto 2.0, Cloudflare Clef, Perplexity Decider and Fastino GLiNER2.5-Decide can be used commercially. So can Chronos-2 and TimesFM-2.5, Amazon’s Mitra-v2 and Inria’s TabICLv2 (BSD-3).
• The middle. Liquid’s d1 open models use the LFM Open License. Mistral’s Timestral reportedly excludes firms above $20M monthly revenue. Qwen3.8 (August 2026) moved to a licence that needs a separate deal for “Model as a Service” businesses, which a decision API is. Read the newest licence on every release.

The builder’s risk is concrete: fine-tuning TabPFN-3.5 or TimesFM-3 for a product is not allowed without a deal. Picking an Apache base on day one saves a rewrite later.

Seven ways of making money showed up in the market:

1. Per input token, output free, in practice per decision. Nobody publishes a flat per-decision price yet; that position is open.
2. Open weights plus a hosted API: Mistral, Perplexity Decider, Cloudflare Clef, Synthefy, Brightband.
3. Open for research, paid for production: Prior Labs, TimesFM-3.
4. Vertical software around a model: Nixtla Enterprise, Archetype’s Newton Agents, Jua; priced by quote.
5. On-prem and air-gapped licences for regulated buyers, the YC sibling request.
6. Data and validation partnerships: the partner supplies data and lab checks, as in CuspAI’s 45-partner foundry or Isomorphic’s and Chai’s pharma deals.
7. Sell the physical output: Orbital sells the materials and hardware it discovers. Highest margin, most capital.

Gross margin is revenue minus the direct cost of serving it, divided by revenue. A decision model only reads its input; it never generates. That reading pass is called prefill, and it costs about 2 × parameters floating-point operations per input token. MFU (model FLOPs utilization) is the share of a GPU’s peak speed you actually get.

The numbers below are an estimate worked out from public prices, not published figures. Assumptions: an H100 on demand at $3.99 an hour (Lambda pricing, read 2026-10-28), a dense peak of 1,979 TFLOPS in FP8, meaning 8-bit numbers (NVIDIA H100), 40% MFU and a GPU busy 100% of the time.

Cost to read 1M input tokens, 2B model, FP8, one rented H100

FLOPs      2 x 2e9 params x 1e6 tokens        = 4.0e15
speed      1,979e12 peak x 0.40 MFU           = 7.9e14 FLOP/s
time       4.0e15 / 7.9e14                    = 5.1 s
cost       5.1 s x $3.99 / 3,600 s            = $0.0056

margin at $0.042   (0.042 - 0.0056) / 0.042   = about 87%

You can rerun it for any size:

# Reproduces the margin estimate above. An estimate, not vendor data.
H100_PER_HOUR = 3.99   # USD, Lambda on-demand, read 2026-10-28
FP8_PEAK = 1979e12     # dense FP8 FLOP/s
MFU = 0.40             # share of peak actually achieved

def cost_per_million_tokens(params: float) -> float:
    """GPU dollars to prefill 1M input tokens."""
    seconds = 2 * params * 1e6 / (FP8_PEAK * MFU)
    return seconds * H100_PER_HOUR / 3600

for size in (0.34e9, 2e9, 9e9, 27e9):
    cost = cost_per_million_tokens(size)
    for price in (0.042, 0.02):
        margin = (price - cost) / price
        print(f"{size / 1e9:>5.2f}B at ${price}/M: margin {margin:.0%}")

How to read it. At 50% real utilization, double every cost; caching, mixture-of-experts backbones and owned hardware lower it. Even so, a $0.02 price on a 27B model looks like a loss leader or a platform play. The independent seller who survives a price war is the one with the smallest model that holds accuracy, and Arize’s results suggest 9B and smaller can come close. Serving for a fraction of a cent shows how to get there.

Two outside benchmarks frame this. Bessemer’s 2025 data put fast-growing “AI Supernovas” at about 25% gross margin, often negative, and steadier “Shooting Stars” at about 60% (Bessemer, 2025-08-13). A small decision model that owns its weights can sit well above both. An app that resells frontier tokens to make decisions sits at the bottom, and is exposed to an endpoint 65-300x cheaper.

Gateways keep their own take near zero. Vercel’s AI Gateway charges no markup on tokens (Vercel docs, 2026-09-08), and OpenRouter has a :floor variant that sorts providers by price (OpenRouter FAQ). A gateway is free distribution, and it also lists every rival next to you, sorted by price.

As of October 2026 the map has four regions. Decision models are copied by every platform with distribution. Tabular and relational models get bought by data-platform owners. Biology and materials soak up the capital because the reward is a wet lab. Time series and weather still host small funded teams, but Google’s free WeatherNext 3 archive (CC BY 4.0) and Apache models from Datadog, Amazon and IBM are commoditizing that layer too.

What nobody knows yet:

• Jev’s size and architecture are undisclosed, so nobody outside TypeSafe knows its true serving margin.
• Clef’s real price ($0.09 or $0.24 for flash) and whether the 27B backbones are dense change the margin maths.
• Deal values for Prior Labs and Kumo are undisclosed or single-source.
• Whether a flat per-decision price with a calibration guarantee would beat per-token pricing is untested.

• YC Requests for Startups — the “Vertical Foundation Models” entry and its siblings, in YC’s own words.
• Arize: decision model benchmark — the first independent multi-vendor test, with calibration, latency and cost.
• InfoWorld on Microsoft Decision-1 — how a frontier lab matched the primitive and the price in under a month.
• SAP News: SAP to acquire Prior Labs — the lab-to-acquisition path in the acquirer’s words.
• Prior Labs docs: Models — the non-commercial terms, read before you build on TabPFN.
• Bessemer: State of AI 2025 — gross-margin benchmarks for AI companies.