ArenaFoundation models
- AnthropicFinance agents shipped inside Microsoft Office; Colossus 1 compute deal finalized.
- Meta AIMuse Spark launched at frontier reasoning tier; four-generation MTIA roadmap with Broadcom published.
- DeepSeekV4 shipped with million-token context; state-backed first round at up to $45B.
- MistralMedium 3.5 released open-weight; Vibe remote agents shipped.
- SpaceXAI$55B Terafab plan filed; Grok Connectors launched.
Anthropic
● MovedFinance agents shipped inside Microsoft Office; Colossus 1 compute deal finalized.Builds the Claude family (Claude 3.x, Opus 4.7). API delivery aligned with AWS and Google Cloud enterprise lanes. Focus on safety controls and deployment policy.
Founded 2021 · San Francisco, US · Private · Backed by Amazon, Google, Spark Capital · ~500 employeesAPI-first
Anthropic
● MovedAbout
Anthropic builds Claude, a foundation model family used in enterprise and developer products. The company is funded primarily by Amazon and Google and operates as a US-based AI lab.
Strategy
Builds the Claude family with inference routed through SpaceXAI's Colossus 1 — 300MW and 220,000 NVIDIA GPUs operated by a rival. Ships finance agent templates running natively inside Microsoft Office. Mozilla credits Mythos with a 13x year-over-year jump in Firefox vulnerability fixes, giving Claude Security a public third-party deployment. Safety posture, enterprise JV, rival-operated compute, and office-suite integration are all running in parallel.
Meta AI
● MovedMuse Spark launched at frontier reasoning tier; four-generation MTIA roadmap with Broadcom published.Builds Llama (open-weight) and Muse Spark (frontier reasoning) under Meta Superintelligence Labs. The MTIA silicon program with Broadcom runs four chip generations through 2027. Open-weight releases shape the baseline; Muse Spark competes at the frontier reasoning tier.
Founded 2004 · Menlo Park, US · Public · ~70,000 employeesOpen-weight
Meta AI
● MovedAbout
Meta AI builds the Llama open-weight model family and Muse Spark, a frontier reasoning model from Meta Superintelligence Labs. Meta is a publicly traded US company; the AI division operates as a product and research unit within the broader business.
Strategy
Launched Muse Spark — a multimodal reasoning model scoring 58% on Humanity's Last Exam, matching the frontier tier held by GPT Pro and Gemini Deep Think. Published a four-generation MTIA silicon roadmap with Broadcom. Meta now has a frontier model, a published multi-year silicon program, and the most widely deployed open-weight model family all running in parallel.
DeepSeek
● MovedV4 shipped with million-token context; state-backed first round at up to $45B.Builds open-weight frontier models with non-NVIDIA hardware targeting. V4 ships million-token context, hybrid attention, and a new RL sandbox. China's state semiconductor fund is in talks to lead a first VC round at up to $45 billion.
Founded 2023 · Hangzhou, China · Private · Backed by High-Flyer · ~200 employeesOpen-weight
DeepSeek
● MovedAbout
DeepSeek is a Chinese AI lab that builds and releases open-weight frontier models. It is backed by the quantitative trading firm High-Flyer and is approaching its first institutional funding round.
Strategy
Ships V4 with million-token context and a new RL sandbox, backed by a state semiconductor fund entering at $45B. No longer dependent on a single founder's balance sheet — this is open-weight frontier work entering a sustained-capital phase. Hugging Face Hub download share for Chinese model lineages reached 41% this cycle.
Mistral
● MovedMedium 3.5 released open-weight; Vibe remote agents shipped.Builds efficient open and commercial models. Positioned as the European frontier lab with a strong open-weight distribution strategy. Medium 3.5 runs on four GPUs under a modified MIT license.
Founded 2023 · Paris, France · Private · Backed by Lightspeed, a16z, Bpifrance · ~150 employeesOpen-weight
Mistral
● MovedAbout
Mistral is a Paris-based AI lab that builds open-weight and commercial models. It is venture-backed and publishes its models for download and through major cloud providers.
Strategy
Released Medium 3.5 — a 128B open-weight model with 256k context and 77.6% on SWE-Bench Verified, runnable on four GPUs under a modified MIT license. For small teams that want a strong coding model on their own hardware without hyperscaler budgets, this is the lowest barrier-to-entry frontier-adjacent option.
SpaceXAI
● Moved$55B Terafab plan filed; Grok Connectors launched.Builds Grok. Distributed through X platform and API. Colossus 1 is the compute foundation — rented to Anthropic. SpaceX filed a $55B Terafab chip plant plan in Texas (expandable to $119B, targeting up to 200GW/year). Grok Connectors launched, opening third-party app integrations.
Founded 2023 · Hawthorne, US · Subsidiary · Part of SpaceX · ~200 employeesAPI-firstInfrastructure
SpaceXAI
● MovedAbout
SpaceXAI (formerly xAI) builds Grok, a foundation model distributed through the X platform and a developer API. It is a wholly owned unit of SpaceX following a corporate dissolution of xAI on May 6, 2026.
Strategy
Compute landlord, chip plant filer, and model distributor simultaneously. Terafab — a $55B Texas chip plant with Tesla and Intel, expandable to $119B — is the biggest single infrastructure commitment disclosed this cycle. Grok Connectors launched, opening third-party app integrations for the first time. Colossus 1 rents to Anthropic while the Grok enterprise channel remains thin.
OpenAI
Builds GPT-4o, o3, o4-mini. Distributes through API and partner channels including Azure and AWS Bedrock. Enterprise channel influence is its primary leverage point.
Founded 2015 · San Francisco, US · Private · Backed by Microsoft, Andreessen Horowitz, Sequoia · ~3,500 employeesAPI-firstDistribution
OpenAI
About
OpenAI builds GPT-4o, o3, and o4-mini and distributes them through its own API and partner channels including Azure and Amazon Bedrock. The company is structured as a for-profit entity controlled by a nonprofit parent and is headquartered in San Francisco.
Strategy
Runs three distribution tracks in parallel: the direct API, Azure/Bedrock channel, and the DeployCo JV with TPG. This multi-channel strategy reduces single-provider lock-in risk for enterprise buyers while maintaining leverage through each lane.
Google DeepMind
Builds the Gemini family (2.x, 3.x). Distributed through Google Cloud Vertex AI and consumer surfaces. TPU hardware advantage directly shapes training economics.
Founded 2010 · London, UK · Subsidiary · Part of Alphabet · ~2,000 employeesPlatformInfrastructure
Google DeepMind
About
Google DeepMind builds Gemini, a foundation model family available through the Gemini API and Google Cloud Vertex AI. It is a division of Alphabet and has access to Google's proprietary TPU accelerator fleet.
Strategy
Keeps a frontier claim active through preview-cadence drops rather than headline launches. Teams adopting Gemini through Vertex AI often end up deeper into Google's stack than planned — BigQuery, Workspace, and Cloud Storage integrations compound the dependency quickly.