Mistral Raises €3 Billion, Qualcomm Signs With AWS, and the US Sides With OpenAI on Copyright
Updated: 6 days ago
Title: Mistral Raises €3 Billion, Qualcomm Signs With AWS, and the US Sides With OpenAI on Copyright
Date: 8 September 2026
Type: Blog
Author: SAASiQ (contact@saasiq.ai)
Word count: 1413 words
Reading time: 6 min
Published: 08-09-2026
Mistral announced a €3 billion funding round led by Samsung on 8 September, reported as the largest equity round ever raised by a European technology company, and on the same day Qualcomm and Amazon Web Services announced a deal for custom AI chips. Cognition, which sells AI coding agents, raised more than $2 billion. In the US the Justice Department backed OpenAI's fair use defence against The New York Times, and Anthropic began offering large customers a way to keep its misuse-detection logs in their own cloud.
Mistral's Series D
Samsung Electronics led the round, with the Scaleup Europe Fund, which EQT manages, and PSG Equity as co-leads. New investors include Advent, funds managed by BlackRock and the Grand Duchy of Luxembourg, and TechCrunch reports that existing backers a16z, Nvidia and Salesforce Ventures also took part. The post-money valuation is above €21 billion. A year ago ASML led Mistral's Series C at a valuation of €11.7 billion.
The dollar figure depends on who did the conversion: Quartz gives $3.5 billion, TechCrunch about $3.58 billion. Mistral says the money will go on compute, infrastructure, commercial growth and international expansion, and TechCrunch reports a target of 1GW of compute in Europe by 2030. Arthur Mensch, the chief executive, told CNBC that Mistral will build and own data centres while also renting capacity. Emmanuel Macron described the round as part of France and South Korea 'building a third way in AI'.
For public bodies and regulated firms in Europe that want an AI supplier based outside the US, the data centre plan is the relevant part. A model provider that owns its compute in Europe can state where its models run without relying on a US cloud provider's commitments, and that is what the data residency clauses in a public-sector tender ask a supplier to confirm.
Qualcomm and AWS
Qualcomm will design custom AI inference chips for AWS across several generations, together with 1.6Tb/s optical connectivity for its data centres. Inference is the work of running a trained model to answer requests, as distinct from training it, and its cost rises with every query a customer sends. Neither company disclosed the chip architecture, the manufacturing process, the foundry or the volumes.
The financial terms are in a warrant Qualcomm issued to Amazon on 3 September and disclosed in an 8-K filing. A warrant gives the holder the right to buy shares at a fixed price. This one covers up to 25 million Qualcomm shares at $161.26 each, about $4.03 billion at most, and runs for ten years. Of those, 3.75 million shares vested when the warrant was issued. The other 21.25 million vest as Amazon makes purchases from Qualcomm, counted up to $60 billion, through to 3 September 2036.
Several outlets reported the $60 billion as the value of the deal. RCR Tech describes it as 'a ceiling for qualifying payments', so Amazon has made no commitment to spend it and it should not be read as Qualcomm revenue or order backlog. Cristiano Amon, Qualcomm's chief executive, said AI infrastructure 'needs advances in both computing and connectivity to deliver more performance efficiently'.
Cognition's Series E
Cognition raised more than $2 billion at a valuation of $48 billion, in a round led by Andreessen Horowitz and Accel, with Founders Fund, General Catalyst and Avenir among the existing investors taking part. The company says its run-rate revenue, current sales projected over a full year, went from $492 million in May to almost $900 million.
Cognition sells agents that write and change software. Its announcement names NVIDIA, GE Aerospace, Citi, Mercedes-Benz and Modal as customers, and puts its view of the job plainly: 'Engineers should operate more like architects and delegate execution to swarms of agents.'
Citi, on that list, is a regulated bank. A bank using coding agents still needs the code they produce to pass the same review, testing and change control as code its own staff write, with a record of which agent made which change, and those controls should be in place before an agent is given write access to a production repository.
The Justice Department and The New York Times
The US Justice Department has filed a statement of interest supporting OpenAI's fair use defence in the copyright case brought by The New York Times, the Times reported on 2 September. The filing was made late on 1 September in the US District Court for the Southern District of New York, where Judge Sidney H. Stein is hearing the case.
The department's position is that training large language models on copyrighted text is transformative and generally fair use, and that the benefits of AI 'far outweigh any competitive harm'. Stanley Woodward Jr., the Associate Attorney General, said: 'AI dominance is critical to promote national security, prosperity, and economic mobility for all Americans.' Graham James, a spokesperson for the Times, said the proposal 'would undermine the sustainability of the human-created content' that society depends on. OpenAI did not comment and Microsoft declined to.
A statement of interest puts the government's view to the court, and the ruling remains Judge Stein's. Summary-judgment motions were due by Friday 4 September. The argument turns on US fair use doctrine, so a ruling would not settle the position under UK copyright law.
Anthropic's Enterprise Frontier Safeguards
Anthropic has launched Enterprise Frontier Safeguards, which Help Net Security covered on 2 September. It lets large customers use Claude under zero data retention, where Anthropic keeps no copy of prompts or responses, while automated checks for misuse still run. The logs those checks produce are written to the customer's own Amazon S3, Azure Blob Storage or Google Cloud Storage, under the customer's encryption keys, access policies and audit logging. No Anthropic employee reviews the data, and flags go straight to the customer's own staff.
It covers Claude Code, Claude Enterprise and the Claude Platform, as well as Claude used through Amazon Bedrock, Google's agent platform and Microsoft Foundry. Anthropic makes no charge for it; the customer pays its cloud provider for the storage. Rollout is phased, with broad availability targeted for the autumn.
Anthropic says it designed the service with more than 100 customers, among them a quarter of the Fortune 100 and every US global systemically important bank, as well as Comcast, KPMG, Mastercard, Salesforce and Visa. Eight members of the Analysis and Resilience Center for Systemic Risk took part. Munish Kumar Sharma, chief information security officer at Wells Fargo, described the result: 'Our logs stay in a Wells-managed environment under Wells-managed keys.'
Misuse monitoring means someone keeps a record of what staff asked the model and what was flagged. Here that record stays in the customer's cloud under the customer's keys, and UK regulated firms and public bodies can ask their other AI suppliers whether they offer the same arrangement.
GPT-6 Astra pricing
OpenAI released GPT-6 Astra, its new flagship model, on 3 September. According to CloudZero's analysis of the price list it costs $10 per million input tokens and $50 per million output tokens, with cached input at $1 and cache writes at $12.50. Batch and Flex processing cost half those rates, Fast mode costs double, and the price changes for requests above 272,000 input tokens. A token is roughly three-quarters of a word.
The context window, the amount of text the model can take in at once, is 1,050,000 tokens, with up to 128,000 tokens of output. The model rolled out to ChatGPT Plus, Pro, Business and Enterprise and to the API over the days after launch, and it is also offered through Azure and AWS Bedrock. OpenAI says it is the first model it has rated 'Critical' for cyber risk under its Preparedness Framework.
Those rates are 2.5 times the promotional price of GPT-5.6 Sol, which stands at $4 per million input tokens and $20 per million output tokens. For work that already runs well on Sol, the gap is large enough to justify testing a sample of real tasks on both models before switching.
For buyers
Anthropic's customer-held logs and Mistral's plan to own data centres in Europe both bear on the same tender question, which is where an AI supplier keeps data and who can look at it. Both are worth putting to incumbent suppliers at the next contract review.
On cost, any budget that assumes GPT-5.6 Sol at its promotional rate needs an end date against it. The Sol promotional rate of $4 and $20 per million tokens is guaranteed until 21 November 2026.
SAASiQ - Intelligent Solutions for SaaS ©
