Harvey's Chinese base model and the repricing of vertical SaaS

On 20 August 2026 Harvey published a short research post announcing Tenet, its first post-trained in-house model for legal work. The base is Kimi K3, the open-weight model from Moonshot AI in Beijing.

Five months earlier the same company had raised $200 million at an $11 billion valuation, co-led by GIC and Sequoia, selling a product that a certain kind of investor kept describing as a form over a database with somebody else's model behind it. Both of those facts are true at the same time, and holding them together is the only honest way to think about what is happening to vertical software.

The generic layer got absorbed. What Harvey did next tells you what was left underneath.

Windsurf was carved into three pieces in seventy-two hours

Windsurf had $82 million in annual recurring revenue and more than 350 enterprise customers. OpenAI agreed to buy it for roughly $3 billion. In July 2025 that deal fell apart. Within days Google paid $2.4 billion for a non-exclusive licence to the technology and for the chief executive, a co-founder and a small group of research leads. Cognition then signed for everything left standing: the product, the brand, the intellectual property and the customer base.

Read those prices separately, because the buyers did. Google paid $2.4 billion and took no revenue at all. Cognition took the revenue and the enterprises and paid an undisclosed number.

One company, two assets, two incompatible logics. Talent and a licence went to a model lab. Distribution and workflow went to an application company.

Nobody bought the wrapper. Everybody bought one of the two things sitting underneath it. That is the cleanest natural experiment anyone has run on the thin-wrapper thesis, and the result is not what the thesis predicted. The wrapper was worth nothing. The wrapper's customers were worth a great deal.

Nobody bought the wrapper. Everybody bought one of the two things sitting underneath it.

The first quarter of 2026 repriced software without reading the filings

SaaS Capital keeps an index of public pure-play SaaS companies. Its median revenue multiple hit decade-plus lows in the first quarter of 2026, and the firm is direct about the cause: markets started pricing AI as an existential threat to the subscription model itself.

The instructive detail sits one level down. SaaS Capital had been splitting its index into higher and lower AI-risk baskets, and through the third quarter of 2025 the two behaved differently, which is what you would see if investors were actually discriminating between business models. In the first quarter of 2026 both baskets fell in lockstep.

That is a sentiment trade, not an analysis.

Meanwhile the underlying companies got duller and healthier. Median growth in the index has come down from more than 30% at the 2021 peak to the low teens, and for the first time in over a decade the median public SaaS company earns an operating profit. Slower, profitable and priced as though it were terminal.

The casualty everybody points at is Chegg, and Chegg is a bad archetype. It sold homework answers. A chatbot gives homework answers away. That is a content arbitrage business meeting a free substitute, which tells you almost nothing about what happens to a claims platform or a hospital scheduling system where the software is entangled with a regulated process.

Harvey stopped renting the model

A company on Harvey's curve has one obvious move available to it. Keep renting frontier models, put every spare dollar into enterprise sales, worry about the model layer in three years when someone else has solved it.

It did the opposite. Tenet is a Kimi K3 base post-trained with Fireworks on public legal data, synthetic data and lawyer-annotated data, using asynchronous reinforcement learning against task environments rather than static evaluations.

The claims in Harvey's own write-up are worth reading slowly. Tenet completes almost twice as many held-out tasks on the firm's internal legal benchmark as the base model does, and 20% more on the contracts subset. More telling than the accuracy: they shaped the reward to prefer trajectories that consume fewer tokens at inference for equivalent performance, and a specialised review-table model now runs at roughly a tenth of the cost per cell.

Cost per cell. That is what a legal AI company at an eleven-figure valuation is optimising in the summer of 2026, and it tells you precisely where the margin fight has relocated.

Once open weights land within touching distance of frontier quality, the generic reasoning layer stops being something you buy and becomes something you configure. What stays scarce is the annotated data nobody else has, the benchmark that reflects what a partner will actually put their name to, and the workflow that delivers the answer to a lawyer who is already logged in and already billing.

Kingdee's interim report reads nothing like an extinction event

I read Kingdee's interim announcement in Chinese the morning it went up. It is a habit worth acquiring if you want to know what enterprise software is doing rather than how American investors feel about it.

First half of 2026: revenue up 13.6% to RMB 3.63 billion, gross margin 67.7%, cloud services now 86% of the total, revenue from AI-native products up 189%, and a swing out of loss into a small attributable profit of RMB 54.5 million.

The Chinese repricing is real. It just runs on different mechanics. Vertical software here never had the American per-seat gross margin cushion in the first place. Deals are implementation-heavy, discounted brutally, and a large share of the money has always sat in the deployment rather than the licence. You cannot raise a seat price that was never really a seat price, so vendors book the AI capability as its own disclosed line instead. Kingdee reports AI-native product revenue separately. Its larger rival Yonyou, loss-making and queuing again for a Hong Kong listing, has spent two years converting the same story into signed order commitments.

Over lunch in Nanshan in June, an implementation partner who does ERP rollouts for mid-sized factories out in Dongguan told me the AI module now sells more easily than the ERP underneath it. The customer buys the demo in an afternoon. Then they find out the migration is still eighteen months of filthy master data, that three of their four production lines write dates differently, and that nobody internally wants to own the mapping. He was not complaining. That gap is his entire business.

There is a second Chinese moat with no technical content whatsoever. The xinchuang rules that favour domestic stacks in government and state-adjacent procurement hand distribution to named vendors by policy. A better model does not get you onto that list.

Brussels wrote a moat that no model can cross

The European Commission awarded a 180 million euro framework in April 2026, running six years, to supply sovereign cloud to the EU institutions. Four providers won it: Post Telecom of Luxembourg with OVHcloud and Clever Cloud, StackIT out of the Schwarz group, Scaleway from Iliad and a Proximus-led consortium including S3NS, Clarence and Mistral.

No American hyperscaler is on that list. Their infrastructure is not worse. Eligibility required SEAL-2 under the Cloud Sovereignty Framework, which scores providers across eight objectives covering jurisdiction, supply chain, operations and technology. There is no engineering answer to a jurisdictional test, because the test is about who ultimately owns you.

This is the only moat in the piece that a better model leaves completely untouched, and it is also the only one a mid-sized vertical software company can acquire without inventing anything. Certification is capital expenditure and patience.

It is not a European peculiarity either. In Lagos this year I sat through a software meeting where the first question was not which model the vendor used. It was where the audit log physically lives, because Nigeria's data protection directive now restricts cross-border transfer by default unless you have a proper instrument in place. The vendors with an answer stayed in the room.

The strongest version of the case against all of this

Salesforce reported its first quarter of fiscal 2027 in May 2026. Revenue of $11.13 billion, up 13%. Agentforce annual recurring revenue past $1 billion. AI and Data 360 ARR combined at $3.4 billion, with half of those bookings coming from existing customers expanding. Current remaining performance obligation of $33.6 billion, up around 14%. Non-GAAP operating margin at a record 34.8%, up 250 basis points. Full-year guidance raised to $45.9 to $46.2 billion.

Then the detail that actually stings. Seven of the ten largest deals in the quarter added net new seats. Agents were supposed to eat seats. In the biggest contracts at the biggest seat-based vendor on earth, the agents turned up with more seats attached to them.

So concede the strong form of the counter-argument. The SaaSpocalypse is a multiple story rather than a revenue story. Sector revenue is still compounding at low double digits. Retention in the deep verticals holds because a hospital does not replace its scheduling system on the strength of a demo. Distribution still takes longer to build than product does. Seat pricing is being hybridised rather than abandoned, a seat floor with consumption stacked on top, which is what nearly every agent SKU currently on sale amounts to. And my three headline examples are all coding tools, the single most model-exposed category in software. Generalising from Cursor to a claims-adjudication platform is a stretch and I know it is.

Here is where I still disagree. The multiple is the signal precisely because it is a statement about terminal value rather than next year's bookings. Two companies both growing 15% are not worth the same money if one owns the data its product runs on and the other rents intelligence from a supplier who has publicly said it wants the application layer. The market cannot yet tell those two apart, which is why it sold them together in the first quarter and bought some of them back in July. It will learn to tell them apart. The tell will be gross margin under agent load, and that number shows up on an income statement roughly two years before it shows up in a narrative.

The base model has a nationality now

SpaceX took an option in April 2026: about $10 billion for a partnership with Cursor, or $60 billion to buy the company outright later in the year. On 16 June it exercised the second, in stock, for a business with reported annualised revenue around $2.6 billion and no frontier model of its own. The largest acquisition of a venture-backed startup on record went to a product that is a workflow and a user base.

Which is the argument, more or less.

The part nobody has priced is the twist at the end of it. Harvey sells to law firms whose entire procurement instinct is confidentiality, and its own model now sits on weights trained in Beijing. The consortium that won a slice of Europe's sovereign cloud contract has Mistral inside it. Cursor belongs to a launch company with government customers.

Six months ago the diligence question in vertical software was which model you use. It is turning into whose law your weights answer to, and there is no line for that on anybody's cap table.

Sources

Harvey, Update on Harvey's Post-Training Effort (Harvey Tenet), 20 August 2026 : https://www.harvey.ai/blog/post-training-update-harvey-tenet

CNBC, Legal AI startup Harvey valued at $11 billion in funding round, March 2026 : https://www.cnbc.com/2026/03/25/legal-ai-startup-harvey-raises-200-million-at-11-billion-valuation.html

CNBC, Cognition to buy Windsurf days after Google poached CEO in $2.4 billion licensing deal, July 2025 : https://www.cnbc.com/2025/07/14/cognition-to-buy-ai-startup-windsurf-days-after-google-poached-ceo.html

CNBC, SpaceX to acquire AI coding startup Cursor for $60 billion, 16 June 2026 : https://www.cnbc.com/2026/06/16/spacex-spcx-cursor-acquisition-ipo.html

SaaS Capital, Four early 2026 SaaS trends : https://www.saas-capital.com/blog-posts/four-early-2026-saas-trends/

Salesforce Q1 fiscal 2027 earnings call transcript, 28 May 2026 : https://www.fool.com/earnings/call-transcripts/2026/05/28/salesforce-crm-q1-2027-earnings-transcript/

European Commission, Commission advances cloud sovereignty through strategic procurement, 17 April 2026 : https://commission.europa.eu/news-and-media/news/commission-advances-cloud-sovereignty-through-strategic-procurement-2026-04-17_en

Kingdee International Software, 2026 Interim Results announcement : https://www.marketscreener.com/news/kingdee-international-software-announces-2026-interim-results-ce7859dad181f02d