Harvey Moves to an Open-Weight Model, Cohere and Aleph Alpha Sign, and Crusoe Raises $3.9bn
Updated: 6 days ago
Title: Harvey Moves to an Open-Weight Model, Cohere and Aleph Alpha Sign, and Crusoe Raises $3.9bn
Date: 22 September 2026
Type: Blog
Author: SAASiQ (contact@saasiq.ai)
Word count: 1577 words
Reading time: 6 min
Published: 22-09-2026
Harvey, the legal AI company, saw its gross margin fall from about 50 per cent to about minus 50 per cent in the first half of 2026 as its model bills rose, and in August it released an in-house model built on a Chinese open-weight base, Bloomberg reported on 21 September. In the same week Cohere and Aleph Alpha signed their merger, Crusoe raised $3.9bn, and Snorkel AI, Heidi Health and Verda announced new funding. British Columbia filed a lawsuit against OpenAI, and Alibaba set out a new AI chip and a larger Qwen roadmap.
Harvey's model costs
Harvey, valued at $15.5bn in its funding round of 9 September, sells AI tools to lawyers. According to Bloomberg, as reported by The Next Web, its gross margin was about 50 per cent at the start of 2026 and about minus 50 per cent by June. The cause was usage: after a March update to Harvey's agents, customer token consumption rose roughly twentyfold, and Harvey was paying for those tokens under OpenAI's and Anthropic's usage-based pricing, which both companies now charge enterprises on top of base subscription fees.
A token is a small unit of text, roughly three-quarters of a word, and model providers bill per million of them. An agent that plans a task, calls tools and checks its own work makes many model calls for each request a user sends, so an update that makes agents more thorough can multiply the tokens behind each piece of work. A software company that pays its model supplier by the token, and prices its own product some other way, carries the difference.
In August Harvey released its first in-house model, post-trained on Kimi K3, an open-weight model from the Chinese lab Moonshot. Open weights means the trained model is published and can be run, and trained further, on infrastructure the company chooses, so the cost becomes compute rather than a per-token fee paid to a lab. Harvey's gross margin turned positive again after that launch and what the report calls other changes to how it uses AI. The coverage says Harvey moved its flagship to the in-house model. It does not say Harvey has stopped using frontier models altogether.
Other companies training their own models
Bloomberg named several other companies moving the same way. Abridge is building a clinical model on Nvidia's open weights. Decagon routes 80 per cent of customer queries through its own models. Legora, based in Stockholm, reached $100m of revenue in 18 months and serves more than 1,200 firms. Ramp, which raised $750m in June, is weighing its first training run, and its co-CEO Karim Atiyeh said of the idea: "It made absolutely no sense a year ago. It's starting to make a lot more sense now."
Cohere and Aleph Alpha sign their merger
Cohere and Aleph Alpha signed a definitive combination agreement on 16 September, formalising a framework first disclosed in April. Reuters puts the combined company's value at about $20bn. It will operate as Cohere, with headquarters in both Toronto and Berlin, and Aleph Alpha's Heidelberg office becomes a research centre. Combined headcount is more than 1,000. Aidan Gomez stays as chief executive, Aleph Alpha's Ilhan Scheer becomes chief operating officer and its co-founder Samuel Weinbach becomes chief research officer.
Aleph Alpha has concentrated on helping governments and businesses deploy AI, and Reuters reports that both firms cited demand for AI that runs inside a customer's own infrastructure and meets local regulation. According to SiliconANGLE, Cohere had $240m of revenue last year and was previously valued at $7bn. Workloads will be hosted on STACKIT, the cloud run by Schwarz Group, which plans an €11bn data centre supporting up to 100,000 GPUs.
Reports differ on the size of Schwarz's own investment. SiliconANGLE says €500m, about $573m, into the upcoming Cohere Series E, and another report gives $600m. SiliconANGLE also mentions a reported investment of up to $3bn from a consortium backed by the Canadian government, which has not been confirmed. The merged company is positioned for public-sector and regulated buyers in Europe and Canada who want an alternative to US hyperscalers' models.
Crusoe and Verda
Crusoe announced the initial closing of an expected $3.9bn Series F on 17 September, at a post-money valuation of $30.9bn. Atreides Management, Mubadala Capital and Valor Equity Partners co-led the round, and Nvidia, GIC and the Qatar Investment Authority were among the other investors. Ten months earlier Crusoe raised $1.38bn at a $10bn valuation, so its value has roughly tripled.
Crusoe reports more than $140bn of total contracted value across the business, more than 6GW of gross contracted capacity and about 1GW already delivered and operating. The money goes into what it calls AI factories, its Crusoe Spark modular data centres and Crusoe Cloud. Chase Lochmiller, the chief executive, said getting there "means controlling the infrastructure from electrons to tokens".
Verda, a European AI cloud with operations in Helsinki, London, Taipei and San Francisco, raised $189m (€163m) on 22 September. Emergence Capital led, with Supermicro and Finnish investors including Varma Mutual Pension Insurance and Tesi. Bloomberg reports a valuation of at least $1bn. Total funding, equity and debt together, is now over $450m, and the annualised revenue run-rate was $165m in July. Its Finnish data centres are live, sites in the rest of Europe, the UK, the US and Asia come online in 2027, and it is aiming for more than 250MW of operations that year.
Snorkel AI and Heidi Health
Snorkel AI raised $350m in a Series E at a $3.5bn valuation, Reuters reported on 22 September, nearly three times the $1.3bn of its Series D. Insight Partners and S32 led the round. Snorkel has moved from selling labelling software to what it calls data-as-a-service: finished datasets and reinforcement-learning environments, the scored practice tasks models are trained on, supplied to frontier labs and corporate customers. Reuters says its annualised revenue run-rate has crossed $350m, up from about $20m a year earlier. TechCrunch gives $375m.
Heidi Health, based in Melbourne, raised $340m in two parts. A $100m Series C was led by Blackbird. The other $240m comes from General Catalyst's Customer Value Fund, which pays for sales and marketing in return for a capped share of the revenue from the new customers, so it does not dilute ownership. MobiHealthNews and BusinessCloud put the valuation at $900m, while SiliconANGLE says it was not disclosed.
According to BusinessCloud, Heidi has 67 staff in Farringdon, London, and plans to double its European headcount within 12 months. It is sole supplier on the NHS England Midlands procurement, which covers 15 acute trusts and 1,239 GP practices and is described as the largest clinical AI procurement of its kind in the NHS. West Hertfordshire, Modality Partnership and One Care are also named as NHS customers, and Heidi supports about 560,000 NHS consultations a week. Its co-founder and chief executive, Dr Thomas Kelly, said: "The ambition was always bigger than writing doctor's notes."
Alibaba's chip and Qwen plans
At its Apsara Conference in Hangzhou on 22 September, Alibaba unveiled the Zhenwu V900, a chip for both training and running AI models. Alibaba claims three times the performance of the Zhenwu M890 it released in May, 216GB of on-package memory and clusters of up to 500,000 chips. The official launch is in the first quarter of 2027.
Qwen 4 is in training, in Max, Flash, Plus and 27B versions. Alibaba projects Qwen 4.5 and Qwen 5 at 5 to 10 trillion parameters, against 2.4 trillion for its current flagship, Qwen 3.8 Max, and has set a target of 20GW of data-centre capacity by 2032. Chinese open-weight models are already inside Western software products, as Harvey's move shows, and Alibaba is now building its own chips as well.
British Columbia's lawsuit against OpenAI
The Province of British Columbia filed a lawsuit against OpenAI and its chief executive, Sam Altman, in US federal court in San Francisco on 21 September. It concerns the shooting at Tumbler Ridge on 10 February 2026, in which eight people were killed, six of them children. The province alleges that OpenAI's safety team flagged the shooter's ChatGPT conversations about gun violence and that the company did not notify law enforcement.
The claim seeks compensation for past and future emergency-response and recovery costs, and an order requiring safeguards that "reliably refuse, terminate or de-escalate" conversations in which users express intent to harm others. In April Altman published a letter saying he was "deeply sorry" OpenAI had not contacted police, and the claim says reforms were not delivered, according to Al Jazeera. A public body suing a model provider to recover its costs is a new kind of liability claim for the industry. These are allegations and have not been tested in court.
For buyers
The Harvey figures show how quickly usage-based model pricing can change a supplier's costs, and a supplier under that pressure may change the model underneath its product between one contract year and the next. Contracts for AI-embedded software should name the model or models in use and where they are hosted, and require notice before either changes. Data protection teams need the same information for their impact assessments.
A supplier moving to a model post-trained from a Chinese open-weight base can have sound commercial reasons, and a public-sector customer will still want to know before it happens, for its own provenance checks.
The Cohere and Aleph Alpha merger is subject to regulatory approval and is expected to close later in 2026.
SAASiQ - Intelligent Solutions for SaaS ©
