For three years, Latham & Watkins has invested to stay at the technological frontier, and the strategy appears to be paying off.
The firm currently runs Harvey across the company, while also leveraging ChatGPT, Claude, and Gemini. And in locked space at a third-party data center, reachable only by Latham staff, the firm keeps several Nvidia servers of its own, where its engineers fine-tune open-weight models on the work the firm does every day.
“We are not hitching our wagon to one particular company,” Rene Mendoza, the firm’s CIO, told the Financial Times. “Nobody can predict where any of this is going, so we have the best optionality.” Legal Cheek called the move a first for BigLaw.
Latham doesn’t train models from scratch. It downloads open-weight models, adapts them to legal work, and wires them into its own software. The firm settled on the direction three years ago, runs Nvidia H200 chips today, and is looking at Blackwell systems next. It won’t say what it paid but Legal Cheek puts a build of this size, hardware plus the specialists to run it, in the tens of millions a year. Not a small chunk of change for an insurance policy, as Harvey still runs firm-wide for research, document analysis, and drafting. Michael Rubin, who leads the firm’s AI work, told Bloomberg Law the buildout was never meant to cut reliance on Harvey. It’s been there as a hedge to ensure they can remain on the frontier, whatever direction the AI industry takes it.
Another great reason to want the optionality came earlier this year when Anthropic launched Claude for Legal in May, with connectors into Harvey, iManage, NetDocuments, Relativity, and Thomson Reuters. Google shipped Gemini Enterprise for Legal in August, and Cleary Gottlieb, Freshfields, Weil, and Williams & Connolly signed on as launch customers. OpenAI hired Ironclad’s co-founder in June to run a legal vertical and is preparing its own legal integrations.
The companies that supply Latham’s models are now selling into Latham’s market. A firm running only rented models is exposed twice: to a vendor’s price and access terms, and to that vendor competing for the same client work.
THE ASSET THAT COMPOUNDS
Legal work is a hard case for the cloud. Privileged material, deal documents, client confidences: the files a firm least wants passing through someone else’s API. Owning the hardware answers that directly, and it does something else too.
Latham has about 100 of its 900 technologists on AI, and it keeps hiring innovation attorneys and AI services attorneys. “No AI company, no technology company will ever be able to replicate what we’re doing because they don’t have our lawyers,” partner Amber Banks told Bloomberg Law.
A firm’s most useful training material is its own casework: the briefs it wrote, the clauses it negotiated, the corrections a senior lawyer made to a junior’s draft. Using a vendor’s model or app out of the box, without integrating your own data into the model training itself, is taking away part of your competitive advantage against other firms using the same providers.
A frontier model arrives knowing nothing about how a particular firm argues a particular kind of case. Latham can take an open-weight model, RLHF using data sets only they have, and improve the models capability for the work they do. Every model generation makes the set more useful, because the firm is applying a better model to better examples. I’m sure they can take their model and also apply it to Harvey (or the legal app of their choice outside of the labs), for the app’s superior workflows.
The technology itself is no longer an edge. For firms like Latham, their advantage will be their relationships and decades of proprietary data.
Pacing the Frontier
Over the weekend, three rivals who rarely agree issued the same warning and offered three different cures. Anthropic’s Dario Amodei called for pacing frontier capability gains. OpenAI’s Sam Altman backed independent evaluators embedded inside the labs. Elon Musk wants peer review by competing labs. Gavin Baker made a useful distinction: the headlines describe a broad slowdown, but the only concrete change so far is third-party evaluation at OpenAI and Anthropic.
While the announcement itself has taken X by storm over the weekend, I was left wondering what does this mean for enterprises deploying AI?
In an environment where yes, the frontier labs are putting out the best product today, many are starting to turn to open-weights. The models aren’t at the frontier but they’re getting very close, and you don’t have to worry about your vendor cutting off your model access or launching a direct competitor into the market.
On top of that, DeepSeek’s own release for v4.1 flash (a model many are saying is nearing or exceeding Opus 5 performance) noted that quality of data mattered just as much if not more than volume of data in this training set. If you’re a well run firm in your industry, who is going to have better data to train models set to do work in your business than you?
Partnerships & Deals
Compute, Energy & Infrastructure
Google committed €13 billion over two years to expand AI and cloud infrastructure in Finland, including data-center capacity and related energy investments.
Positron AI raised an $875 million Series C at a $5 billion valuation to bring its Asimov inference silicon and Titan systems to market.
TAR raised a $120 million Series A led by Spark Capital at a $1 billion post-money valuation. The company builds modular, off-grid renewable power and battery systems for AI data centers.
Qualcomm and Amazon entered a multi-generation collaboration on customized silicon for AWS AI inference and optical connectivity for data-center networks.
Oracle booked more than $30 billion of new AI cloud contracts in its latest quarter, lifting remaining performance obligations to $664 billion. It also reported 850 megawatts of added capacity, more than 300,000 additional GPUs, and a completed $20 billion equity sale.
Palantir named Nebius its preferred sovereign AI infrastructure partner. Nebius compute and inference endpoints will sit inside Palantir’s enterprise perimeter so eligible customers can control their compute, data, and models.
Data, Knowledge & Retrieval
OpenAI signed content partnerships with BCCL, publisher of The Times of India and The Economic Times, and the Indian Express Group. ChatGPT can surface attributed summaries and excerpts from live and archived reporting across seven languages.
OpenAI launched ChatGPT for Financial Services after design work with Morgan Stanley and Evercore. Premium data from Daloopa, PitchBook, and LSEG News is indexed and hosted by OpenAI with source-level citations.
Box partnered with OpenAI so users can browse, access, and work with governed Box content directly inside ChatGPT.
Versos AI launched an agentic curation system built with NVIDIA NeMo and LangChain that turns natural-language requirements into structured, rights-cleared video training datasets while preserving ownership and provenance records.
Universal Music Group and ElevenLabs signed a multi-year licensing and product-development agreement. Their first product will let fans create with music from participating artists, with additional AI audio tools planned for artists and songwriters.
Models, Training & Developer Tools
Mistral AI raised €3 billion in a Series D at a post-money valuation above €21 billion. Samsung led the financing and paired its investment with plans to deploy Mistral models inside semiconductor operations.
Harvey raised $550 million at a $15.5 billion valuation to expand its legal and professional-services AI platform.
Harvey acquired Guardrails AI, an agent-security and simulation company whose tools test where AI agents depart from intended behavior. The team will join Harvey’s product and engineering organization.
Cloudera and Mistral AI partnered to run and customize Mistral models over governed enterprise data across cloud, on-premises, edge, sovereign, and air-gapped environments.
Meta agreed to acquire Swedish agent startup Stilla.ai to strengthen Meta Business Agent across WhatsApp, Messenger, and Instagram.
Enterprise Deployment & Distribution
Accenture and Google Cloud created the Accenture Gemini Enterprise Business Group. The companies plan a 1,000-person forward-deployed engineering workforce to implement agentic AI and data projects.
OpenAI and the U.S. General Services Administration signed a 27-month OneGov agreement offering eligible federal, state, local, and tribal organizations a zero-dollar license fee and 50 percent off usage. Eligibility extends across a public-sector workforce of approximately 23 million.
Amazon Ads partnered with OpenAI to let selected U.S. advertisers extend Amazon Ads campaigns into a ChatGPT Ads pilot.
Palantir and NVIDIA introduced a sovereign AI stack for critical supply chains, combining Palantir’s ontology and software with customized NVIDIA Nemotron models. The first deployment is inside NVIDIA’s own supply chain.
Cisco, Palantir, and NVIDIA are integrating Palantir’s cybersecurity ontology with Cisco’s Secure AI Factory to give regulated organizations an AI architecture they can operate under their own controls.
Fujitsu expanded its partnership with Palantir, signing a new agreement for AIP and Foundry and becoming a Global Forward Deployed Engineering partner for sales, use-case design, and implementation.
Avid and Google Cloud expanded their partnership with a browser-based Media Composer and new Gemini Enterprise and BigQuery integrations for media search, editing, and production workflows.
Morgan State University and Google Public Sector partnered on an AI research campus with high-performance GPU infrastructure, sovereign-AI governance, cybersecurity tools, and a Google-focused Center of Excellence.
Meta launched Muse with Stripe Link purchase protection and one-time payment cards. Shop Pay and 1Password integrations are planned as additional commerce and credential layers.
Napster and Kameha Ventures formed a strategic partnership to develop locally guided multimodal-agent deployments for government and private-sector customers in the United Arab Emirates.
Industry Applications & Workflows
Sony Semiconductor Solutions and Aramco signed a non-binding memorandum to combine Sony sensing and edge-AI technology with Aramco’s industrial data, infrastructure, and operating environments.
UNDP and NEC signed an agreement to use AI and environmental data for nature conservation, climate resilience, supply chains, and assessment of digital infrastructure including data centers.
Physical AI & Robotics
Maven Robotics launched with a $100 million Series A to build general-purpose robotic systems for material handling and assembly in logistics and manufacturing.
Vecna Robotics raised $31 million led by Unless to scale autonomous forklifts, tuggers, case-picking systems, and new dock automation capabilities.
Swarmer signed a definitive agreement to acquire Ukrainian unmanned-ground-vehicle maker Ratel Robotics for cash and stock worth up to $224 million if earnout milestones are met.
NEURA Robotics and SECO partnered on compute modules for NEURA’s cognitive robots, including the 4NE1 humanoid, and on real-world industrial data and automation for semiconductor manufacturing.
JOYX and Hitch Interactive partnered to combine verified human and robot demonstration data with standardized robot interfaces and real-world deployment environments.
Science & Healthcare
Owkin licensed its K Pro AI scientist and multimodal patient data from the MOSAIC network to Servier for oncology research. This is a new September 11 agreement, separate from Owkin’s earlier Boehringer Ingelheim license.
Tempus AI received an award of up to $9.5 million from ARPA-H to develop and clinically validate an autonomous AI agent for heart-failure care using clinical and consumer-generated data.
IBM and NASA released an open-source Lunar Foundation Model and a machine-learning-ready dataset combining more than 30 aligned layers from nine instruments across four lunar missions.