(Not Just Words)
Do you even remember the world before 2022? It’s not even 5years but the lives of all of us have changed in all possible dimensions! Also something really shifted in particular in 2025 and 2026. The money, talent, even our daily conversation (at least in Silicon Valley) — all of it moved past language models. Toward something harder and, honestly, more interesting: AI that can reason about the physical world.
Call it what you want. World models. Spatial intelligence. Physical AI. The labels are messy and half the founders have their own favorites 🙂 But the underlying bet is the same: that the next frontier isn’t making text generation smarter. It’s teaching AI what happens when you push a glass off a table.
I’ve been following this closely, and I want to give you a real overview of who’s building what, how they think about it, and what I actually believe. This is going to be a long one. Get a coffee! ☕️☕️
I’ll tell you what draws me to this. I’m an optimist. I genuinely believe technology, when it’s built thoughtfully, makes people’s lives better. I grew up watching the internet reshape how people connect and learn. I’ve spent years backing companies trying to solve hard physical problems — robotics, industrial automation, space, energy. And what I’ve seen over and over again is that the missing piece isn’t the hardware. The hardware has gotten incredible. The missing piece is the brain.
A robot that can repeat the same motion in a controlled factory is impressive but limited. A robot that can walk into your kitchen, figure out that something’s about to fall, and catch it — that’s useful. That’s the difference between a very expensive machine and something that genuinely changes how people live and work.
That’s what these labs are trying to build. And I find it absolutely thrilling.
Thinking Machines Lab — The One That Actually Inspired This Post
Founded by: Mira Murati (CEO, former OpenAI CTO), John Schulman (Chief Scientist, OpenAI co-founder), Lilian Weng, Andrew Tulloch, Luke Metz — essentially an OpenAI reunion (February 2025)
Funding: $2 billion seed at $12 billion valuation (July 2025, a16z-led), then a $5 billion Series B at $50 billion valuation (March 2026, a16z and Sequoia). Nvidia is both an investor and a strategic partner — the company will deploy at least one gigawatt of Nvidia’s Vera Rubin systems for frontier model training. Also signed a multibillion-dollar Google Cloud deal.
Investors: Andreessen Horowitz, Sequoia, AMD, Cisco, Nvidia
The model: Two things shipping right now!
First, Inkling — dropped literally a few days ago (July 15, 2026). Their first foundation model, built entirely from scratch. 975 billion parameters in a mixture-of-experts architecture, drawing on roughly 41 billion active parameters per prompt to keep costs low. Trained on 45 trillion tokens across text, image, audio, and video, with outputs currently text-only. It’s open-weight, meaning developers can download and customize the full model without licensing fees. They’re clear it’s not the strongest model available — the bet is that enterprises care less about the smartest general-purpose model than one they can make their own.
Second, Interaction Models — previewed in May 2026. Full-duplex AI that processes input and generates responses simultaneously rather than sequentially, enabling the AI to respond mid-conversation in a manner closer to a natural phone call than a turn-based text exchange. Their initial model, TML-Interaction-Small, claims a response latency of 0.40 seconds. While every other voice model waits for you to finish speaking, Thinking Machines’ approach interrupts and adds context as a human would. It’s a genuinely different interaction paradigm, not just a faster chatbot.
Tinker, their developer platform for fine-tuning open-source models, was the first product out of the gate in October 2025.
🤔 What I think: This company is the reason I wrote this post 😉 I’ve been following Mira Murati since she was CTO at OpenAI, and her thesis is one I find genuinely compelling: that the future of AI isn’t one locked-down frontier model that everyone rents from one of three labs, but AI that’s customizable, accessible, and built to collaborate with humans rather than just respond to them.
The Inkling release is a meaningful moment. After a year of fundraising headlines and no model, they shipped something real. The open-weight commitment is a deliberate statement — and given Murati’s history watching OpenAI retreat from openness on GPT-2, it feels personal.
The interaction model concept is the one that really gets me. The idea that AI should interrupt you (as I sometimes do 👀), add context, respond in 400 milliseconds like a person would — that’s a fundamentally different design philosophy from everything else in this list. Whether it becomes a real product category or gets absorbed into existing assistants is the open question. I think it will become real.
Also, here is a fun fact 🙂 Zuckerberg tried to acquire Thinking Machines and, when Murati said no, reportedly tried to poach her employees. She fired her own CTO for unethical conduct. The dynamics level here is off the charts, and I mean that respectfully.
AMI Labs — The One That Makes Europe Matter
Founded by: Yann LeCun (Chief Scientist, Meta AI, until 2025) and Alexandre LeBrun (previously CEO of Nabla, Meta’s conversational AI lead)
Funding: $1.03 billion seed at a $3.5 billion pre-money valuation (March 2026) — the largest seed round in European history
Investors: SBVA (SoftBank Ventures Asia), and others not yet fully disclosed
The model: World models, specifically designed to reason about the physical world. LeCun has been arguing for years that LLMs hit a ceiling because they operate purely in the space of text — they can describe the world but they don’t understand it. His thesis: you need a system that can predict the next state of a real environment, not just the next word in a sequence.
Think of it this way. If you nudge a glass toward the edge of a table, you already know it’s going to fall. You don’t compute this — you just know, based on years of physical experience. LLMs don’t know this. They’ve read thousands of descriptions of glasses falling, but they’ve never experienced physics. AMI wants to build AI that has.
🤔 What I think: I love the intellectual honesty here. LeBrun refuses to use the word “AGI” or “superintelligence” — he thinks both are meaningless. His quote from a recent TechCrunch interview: “What is superintelligence? I don’t know. It’s not a very useful word.” That kind of groundedness is rare in a space where everyone is racing to make the biggest claim. The whole company feels like it was built by people who’ve shipped real products (LeBrun ran an AI health startup before this) and are tired of hype.
The Asia strategy is smart too. They need real-world industrial partners to train on physical data — robots, factories, manufacturing lines. Korea has all of that plus a government actively throwing money at AI infrastructure. Makes total sense.
No product yet. No timeline either — “We’ll make a surprise when we’re ready,” LeBrun said. I don’t know whether to find that charming or alarming. Probably both!😉Check out this cool LEx’s episode with LeBrun.
World Labs — The Other Big World Model Bet
Founded by: Fei-Fei Li (Stanford professor, former Chief AI Scientist at Google Cloud, creator of ImageNet)
Funding: World Labs raised $230 million seed (September 2024), then $1 billion (February 2026) at a ~$5.4 billion valuation. Total: $1.23 billion.
Investors: Andreessen Horowitz (seed), then Autodesk ($200M anchor), AMD, Nvidia, Fidelity, Emerson Collective
The model: “Spatial intelligence” — Fei-Fei’s framing for AI that can perceive, generate, and interact with the 3D world. Their first product, Marble (launched November 2025), takes text, images, or video and generates persistent, navigable 3D environments. You can walk through them, edit them, export them.
The target markets right now are gaming, VFX, and virtual reality. But the real long-term prize — which she’s very open about — is robotics.
🤔 What I think: Fei-Fei Li is one of the most credible people in all of AI. She created ImageNet, which is arguably the dataset that kicked off the modern deep learning era. She’s been thinking about how computers see and understand the physical world for 20+ years. The Marble product is genuinely interesting as a creative tool, but the robotics thesis is what I’m watching.
What I find particularly compelling is who backed this. Autodesk putting in $200M isn’t just a financial bet — they’re a design software company that wants their users to be able to generate 3D worlds as easily as they write a sentence. That’s a clear commercial path, not just research.
The Nvidia and AMD investments are also telling. Both chip companies are betting that world models drive a new wave of compute demand. They’re not wrong.
Ah, and the most recent fun fact here as well: World Labs just acquired SceniX, a robotics company helping robots learn in simulation and improve in the real world! Curious how it will develop!
OpenAI — The Giant Everyone’s Chasing
Founded by: Sam Altman (CEO), Greg Brockman, Ilya Sutskever, and others (2015). Elon Musk co-founded and funded it early, left the board in 2018, then started xAI to compete against them. That story keeps getting stranger.
Funding: $122 billion round (March 2026) at $852 billion valuation. Total funding well over $200 billion when you include Microsoft’s multi-year compute commitments.
Investors: Microsoft (~$13 billion committed plus Azure infrastructure), SoftBank, Abu Dhabi’s MGX, Fidelity, T. Rowe Price, and basically every major institutional investor who wanted exposure.
The models: OpenAI has been shipping fast. The current flagship family is GPT-5.6, released July 9, 2026, and it comes in three tiers:
- Sol — the workhorse. Leads the Artificial Analysis Coding Agent Index at 80, which OpenAI says is 2.8 points above Anthropic’s Claude Fable 5 on that specific benchmark. Sam Altman claims Sol is 54% more token-efficient on coding tasks than previous versions.
- Terra — the balanced middle tier. Good for everyday enterprise work.
- Luna — budget-friendly, fast. Priced at $1 input / $6 output per million tokens.
All three have a 1.05M-token context window. GPT-5.6 also introduces Programmatic Tool Calling, which runs model-written JavaScript in an isolated V8 runtime, and an “ultra” mode that runs four agents in parallel — bumping Terminal-Bench 2.1 from 88.8% to 91.9%.
OpenAI is calling it their strongest cybersecurity model yet.
To put GPT-5.6 in context: the GPT-5 family has been moving at a pace most people can’t track. GPT-5.1 (November 2025), GPT-5.2 (December 2025, apparently accelerated after a “Code Red” memo when Gemini 3 Pro launched), GPT-5.3, 5.4 (March 2026), 5.5 “Spud” (April 2026), and now 5.6. That’s six major releases in eight months. Whatever else you think about OpenAI, they’re not standing still.
The honest benchmark picture: Sol leads on coding agent tasks. On SWE-Bench Pro, though, Sol scores 64.6% against Claude Mythos 5’s 80.3%. That’s a 15-point gap on software engineering. So the “who’s winning” answer genuinely depends on what you’re measuring.
🤔 What I think: OpenAI is extraordinary at one thing above all others: shipping. Whatever you think about the research direction or the governance drama, their ability to turn frontier research into products that hundreds of millions of people use is rare. ChatGPT is still the fastest-adopted technology in history. That’s a huge achievement!.
What gives me pause is everything happening around the edges. The Apple lawsuit alleging trade secret theft broke in July 2026. The 2025 conversion from nonprofit to public benefit corporation felt messy and lost them some key safety researchers. The original mission was to build AI that benefits all of humanity. At an $852 billion valuation with Microsoft as your biggest infrastructure partner, the incentives get complicated.
I use their products. I think GPT-5.6 Sol is genuinely impressive on coding tasks. I just think the gap between their branding (”safe AGI for humanity”) and their actual decisions is wider than it used to be but I highly respect Sam and how he managed to bring this life changing product to the market. The amount of good it’s already made is invaluable!
Anthropic — The Safety-First Lab
Founded by: Dario Amodei (CEO), Daniela Amodei (President), and seven other ex-OpenAI researchers (2021). They left specifically because they thought safety was being deprioritized as OpenAI scaled. That original conviction has shaped everything they’ve built since.
Funding: $65 billion Series H (May 2026) at a $965 billion valuation, making it briefly more valuable than OpenAI. Total raised: over $132 billion. Led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia. Amazon has committed up to $33 billion. Google up to $40 billion. When the two largest cloud providers are your biggest backers, you are not a scrappy startup anymore.
Investors: Altimeter, Dragoneer, Greenoaks, Sequoia, Amazon, Google, GIC, Coatue, Blackstone, Fidelity, Lightspeed, and a long list of sovereign wealth funds and institutional investors.
The models: The latest publicly available model is Claude Fable 5, released June 9, 2026 — the first publicly available Mythos-class model, a tier that sits above the Opus line. The model naming here tells a story. Fable comes from the Latin fabula, “that which is told,” akin to the Greek mythos. The safeguards are what distinguish Fable from Mythos. In other words: same underlying model, different guardrails.
Fable 5 and its restricted sibling Claude Mythos 5 share the same architecture — 1 million token context window, up to 128K output tokens, priced at $10/$50 per million input/output tokens. In high-risk areas like cybersecurity, biology, and chemistry, Fable 5 blocks responses and falls back to Claude Opus 4.8.
The launch was anything but smooth. Three days after release, on June 12, the US government applied export controls to both Fable 5 and Mythos 5, requiring Anthropic to restrict access to foreign nationals. Because the order took effect immediately with no way to verify nationality in real time, they suspended access for all users globally. Controls were lifted June 30 and access restored July 1.
Then, Anthropic extended free access to Fable 5 through July 19 — for the second time in a week — in what reads clearly as a competitive response to OpenAI’s GPT-5.6 Sol launch. The model wars are very real and very visible right now.
On benchmarks: Fable 5 scores roughly 80% on SWE-Bench Pro. Sol scores 64.6% on that same benchmark, a 15-point gap in Fable’s favor on software engineering. Sol leads on the Coding Agent Index. Fable leads on SWE-Bench. Depending on what you’re building, the answer to “which is better” is genuinely different.
🤔 What I think: I have a soft spot for Anthropic — and not just because I applied for roles there (full disclosure). They’re the lab most honestly grappling with what it means to build something this powerful. Constitutional AI, the Interpretability team, the decision to release Fable with hard safety limits rather than strip them for benchmark performance — these are real choices that cost them something.
The export controls drama is genuinely interesting and I don’t think it’s getting enough analysis. The US government blocking a frontier model release three days after launch, then lifting it two weeks later — that’s a new dynamic in AI that nobody has a playbook for yet. Anthropic’s response (working closely with the government rather than fighting it) tells you something about how they see their role.
At $965 billion and not yet cash-flow positive, the financial picture is still a bet on future dominance. I think it’s a reasonable bet. I also think the Amazon and Google concentration is something to watch.
Mistral AI — Europe’s Answer
Founded by: Arthur Mensch (CEO, ex-DeepMind), Guillaume Lample (Chief Scientist, ex-Meta AI), Timothée Lacroix (CTO, ex-Meta AI) — all in their early 30s when they started (2023)
Funding: $3 billion+ total. Seed: €105M (Lightspeed). Series A: €385M (a16z). Series C: $2 billion (September 2025, ASML anchor with 11% stake at $13.8B valuation). Latest: $830M debt round (March 2026).
Investors: Lightspeed, Andreessen Horowitz, ASML (largest shareholder), BNP Paribas, Bpifrance, HSBC, Eric Schmidt, and many European institutions
The model: Open-weight LLMs, primarily. Mistral’s identity is built on openness and efficiency. They release model weights publicly (Mistral 7B, Mistral Large, Codestral, and others) under Apache 2.0 licenses, which means anyone can run them. Their models are consistently smaller and cheaper to run than American equivalents at similar capability levels. Their latest lineup includes Mistral Small 4, Medium 3.5, and Large 3.
They also launched Workflows (April 2026) for enterprise orchestration and Forge (March 2026) for companies that want to train and own their own proprietary models on-premises. They acquired Emmi AI (May 2026) to push into industrial and physics-AI applications.
🤔 What I think: Mistral is, to me, the most interesting company in European tech. Three ex-DeepMind/Meta researchers with zero finished products raised $100 million purely on their reputations and a whitepaper. Then they shipped. Mistral 7B was a genuine shock when it came out — a model that small shouldn’t have been that good.
The French government’s involvement is fascinating from a geopolitical angle. Macron was on stage with Jensen Huang announcing Mistral Compute. Europe desperately wants an AI champion that isn’t beholden to American infrastructure, and Mistral is the closest thing to it. The ASML partnership is particularly clever — the Dutch semiconductor company that makes the machines every chip fab in the world depends on is now the biggest Mistral shareholder. That’s a serious moat in a world of export controls.
The honest weakness: they’re still playing catch-up on frontier capability benchmarks against GPT-4o and Claude. Great for enterprise cost efficiency. Harder to argue for if you need the absolute best model.
DeepSeek — The One That Changed Everything
Founded by: Liang Wenfeng (CEO), who also co-founded and runs High-Flyer, an $8 billion Chinese quantitative hedge fund (2023)
Funding: Entirely self-funded from High-Flyer profits until June 2026, when they raised $7.4 billion (their first-ever external round) at a $50 billion valuation. Led by Tencent and CATL (the battery giant). China’s National AI fund got the only investor vote.
The model: Open-weight LLMs and reasoning models. Their R1 model (January 2025) was the moment that genuinely shocked the entire AI industry. R1 matched or exceeded frontier American models on most benchmarks, reportedly trained for around $6 million. For context, GPT-4 reportedly cost over $100 million to train. This wasn’t a small efficiency improvement — it was an order of magnitude. US tech stocks dropped nearly $1 trillion in a single session when it came out.
Their V3 and V4 models continue to compete at the frontier. V4 offers 1M context window, open weights, and aggressively cheap API pricing.
🤔 What I think: Liang Wenfeng is one of the most interesting people in tech, period. He started stockpiling Nvidia GPUs in 2019 before most people outside semiconductor circles were paying attention. He funded an AI lab out of hedge fund profits with zero external investors for years. And then he shipped something that forced an entire industry to rethink its assumptions about what AI development costs.
The geopolitical angle is real and can’t be ignored. The June 2026 funding structure gives China’s state AI fund the only voting rights on the investor side. That’s a governance reality every company considering deploying DeepSeek needs to sit with. Their models are open-weight and widely used — but the governance sits in a very specific place.
That tension between “incredibly good and open” and “funded by a hedge fund with Chinese state investors holding the only votes” is genuinely complicated. I don’t have a clean take. I think both things are true at once.
Also: Liang is now personally worth $36 billion, more than Dario Amodei and Greg Brockman. A quant from Hangzhou is the world’s richest AI founder. Wild.
/🤝 Passionate about HardTech? Have comments, want to share some thoughts or ask about sponsorships? Let’s connect!/
/👉 Want to feature your company in HardTech Reads? Apply Here/
Kimi / Moonshot AI — China’s Comeback Story
Every lab in this series has a moment where ambition gets tested by reality. For Moonshot AI, that moment came twice in less than a year — first when DeepSeek’s efficient, cheap, open-weight models pulled Kimi from third to seventh in China’s monthly active-user rankings, and again this month, when Moonshot answered with Kimi K3.
Founded by: Yang Zhilin and a group of Tsinghua classmates — who reportedly named the company after Pink Floyd’s The Dark Side of the Moon on its 50th anniversary — Moonshot built its early reputation on Kimi, an assistant known for handling unusually long documents. When DeepSeek arrived with cheaper, open, competitive models, Moonshot could have retrenched. Instead, Yang pushed the company toward a more aggressive open strategy: bigger models, multimodal capability, agentic orchestration, and public weights, iterating from K2 through K2.5 and K2.6 in rapid succession.
Kimi K3, released in mid-July, is where that strategy fully pays off. At 2.8 trillion parameters, it is the largest open-weight model ever released and the first to approach the 3-trillion-parameter class, with a 1-million-token context window built for long-horizon coding, reasoning, and knowledge work. Moonshot made the full weights publicly downloadable this week, letting any organization with sufficient compute run, fine-tune, and deploy it regardless of jurisdiction.
The performance story is genuinely competitive rather than purely promotional. Moonshot claims K3 performs on par with Anthropic’s Claude Fable 5 and beats Claude Opus 4.8, GPT-5.6 Sol, and GPT-5.5 on GPU kernel optimization benchmarks. Independent third parties lend some support: Arena.ai’s blind evaluations ranked K3 first among developers for front-end coding tasks, and Vals AI placed it second overall — behind only Claude Fable 5. Moonshot itself is candid that K3 still trails the strongest proprietary systems on overall performance, which is a more honest framing than most launch announcements offer.
The market treated the release as a genuine competitive event, not just a press cycle. Shares of rival Chinese labs Z.ai and MiniMax fell as much as 30% and 16% respectively in Hong Kong trading, and even Alibaba dipped. Moonshot had to pause new subscriptions within days of launch because demand exceeded capacity — a rare problem for an AI lab to have. The underlying business is scaling accordingly: daily revenue has grown roughly sixfold since K3’s debut, annualized recurring revenue reached $300 million in June (up from $200 million in April), and Moonshot is reportedly raising a new round at a $50 billion valuation ahead of a possible Hong Kong IPO.
None of this comes without friction. Washington has taken notice: the White House Office of Science and Technology Policy has accused Moonshot of training K3 on restricted Nvidia chips and of large-scale distillation against U.S. models, including Anthropic’s. Moonshot has not responded publicly to the allegations. There’s also a practical tension built into the release itself — a 2.8-trillion-parameter open-weight model is a genuine gift to the research community, but running it locally could require hundreds of thousands of dollars in hardware, putting true on-premises deployment out of reach for all but the largest organizations. None of the major hyperscalers (AWS Bedrock, Azure Foundry, Google Vertex AI) had integrated K3 at launch, so cloud access — with its own cost, data-residency, and vendor questions — remains the realistic path for most users.
🤔 What I think: Taken together, Kimi K3 is less a single product launch than a marker of how fast the competitive frontier is moving inside China itself. Moonshot wasn’t just answering DeepSeek — it was answering Z.ai, MiniMax, and every other lab racing toward lower cost, greater openness, and larger context windows, while also drawing direct geopolitical scrutiny from Washington in the process. A company that looked like it was losing its position less than two years ago is now setting the pace — with all the commercial upside and political exposure that comes with it.
Google DeepMind — The Research Lab That Ships
Founded by: Demis Hassabis (CEO, chess prodigy, neuroscientist, co-creator of AlphaGo) — DeepMind was founded in London in 2010 and acquired by Google in 2014 for $500M. Merged with Google Brain in 2023 to form Google DeepMind.
Funding: Backed by Google/Alphabet entirely. No external funding needed.
The model: Gemini is the main product family (Gemini 2.0 Flash, Gemini Ultra). But DeepMind’s research goes much further: AlphaFold (protein structure prediction), AlphaStar (StarCraft), Genie (world model for games and 3D environments), and robotics work through projects like RT-2 and Gemini Robotics.
The Genie family is their world model play — generating and simulating 3D environments from visual data. It’s research-stage but technically impressive.
🤔 What I think: DeepMind is, by any honest measure, the best AI research lab in the world on breadth. AlphaFold alone earned a Nobel Prize (chemistry, 2024). The work they’ve done across protein folding, game-playing, mathematics, and robotics is unmatched.
Gemini as a product is good but hasn’t had the breakout moment that ChatGPT or Claude have had. There’s a gap between DeepMind’s research quality and Google’s product execution speed that feels structural. Google has been trying to fix this for three years. The jury’s still out.
The robotics work through Gemini Robotics is what I watch most closely. They have the research depth to win here if they can translate it to product.
Meta AI — Open Source at Scale
Founded by: Facebook/Meta internal. Key figure: Yann LeCun as Chief AI Scientist (until 2025 when he left to found AMI Labs). Currently led by LeCun’s successor.
Funding: Zuckerberg announced $60 billion+ in AI capex for 2025. Meta AI is internal — no separate funding needed.
The model: Llama family (Llama 3, Llama 3.1, Llama 4). Fully open-weight. Meta’s bet is that open-source AI strengthens their platform businesses while also being the right thing to do for the field. They’ve been consistent on this for longer than most.
🤔 What I think: Meta’s contribution to open-source AI is genuinely significant and I don’t think it gets enough credit. Every small lab, every developer, every researcher who can’t afford OpenAI API costs runs on Llama. The ecosystem that’s grown around it is real.
The weird thing about Meta AI is that it’s a $1.2 trillion company’s internal R&D division, and it moves faster than most startups. When Llama 3 dropped, it was competitive with GPT-4 on many benchmarks within days of release. That speed is rare.
LeCun’s departure stings a bit, honestly. He was their most intellectually distinctive voice. But the Llama releases keep coming.
xAI — Elon’s Lab
Founded by: Elon Musk (CEO), with researchers from DeepMind, OpenAI, and others (2023)
Funding: $6 billion (2024), then $6 billion more (early 2025), then $20 billion Series C (December 2025). Valuation around $50 billion. Total: over $30 billion raised.
Investors: Sequoia, Andreessen Horowitz, Fidelity, Valor Equity Partners, and various Musk-adjacent investors
The model: Grok (Grok-2, Grok-3). Runs on X (Twitter) and via API. Has access to real-time X data, which is a genuine differentiator for current-events queries. Grok-3 scored competitively on frontier benchmarks.
🤔 What I think: I’ll be honest: xAI is hard to evaluate clearly because the company and the founder are inseparable in a way that affects every decision. The X integration is smart — real-time data access is something no other model has at that scale. Grok-3 is legitimately good. It’s really precise and trustworthy with verifying information.
But the governance, the brand, and the distraction of everything else Musk is simultaneously running makes it hard to assess as a standalone AI company. I don’t personally use it much. I know people who do and like it.
A Few Others Worth Watching
Cohere (Toronto) — focused on enterprise. Aidan Gomez (ex-Google Brain, co-author of the original Transformer paper) is CEO. Raised over $1 billion. Not consumer-facing — their whole product is API access for businesses that want to run models on their own data.
Inflection AI / Pi — Reid Hoffman and Mustafa Suleyman’s company. Suleyman left to become CEO of Microsoft AI. The company has largely wound down its consumer product.
Aleph Alpha (Heidelberg, Germany) — the European enterprise AI play after Mistral. Less known outside Europe but has strong German industrial partnerships.
Zhipu AI / Z.ai — Chinese, research-heavy, recently IPO’d on the Hong Kong Stock Exchange. Their GLM-5.2 model targets long-horizon coding and enterprise tasks.
MiniMax (China) — also recently IPO’d in Hong Kong. Strong on video and audio generation.
How They’re Different: A Quick Map
If you’ve made it this far, here’s how I actually think about these labs relative to each other:
The LLM-first labs (OpenAI, Anthropic, Mistral, DeepSeek, Kimi) are building on the same basic insight: scale language models, add reasoning, make them useful. They differ enormously on openness, safety philosophy, and geography — but the core technology bet is similar.
The world model labs (AMI, World Labs, DeepMind’s Genie work) are making a different claim: that language alone isn’t enough, and the next chapter requires AI that understands physics, space, and cause-and-effect in the physical world.
I think both things will be true for a long time. LLMs are going to keep getting better and keep being useful. And world models are going to start proving themselves in robotics and spatial applications. They’re not competing — they’re building different parts of the same future.
What I find exciting is that we don’t know which labs win, and I don;t think there will be one only. Google DeepMind has the best research. OpenAI has the best distribution. Anthropic has the best safety work. DeepSeek shocked everyone on efficiency. Mistral is doing it all with a fraction of the headcount. AMI and World Labs haven’t shipped anything yet.
The next three years are going to be genuinely unpredictable. And that’s the most exciting thing I can say about a field.
Thanks for reading. If you found this useful, share it with someone who’s trying to make sense of the AI landscape. And if you have opinions on any of these labs — especially strong disagreements — I’d genuinely love to hear them. Hit reply.
P ✌️
BONUS Lab — Have you heard of Prentis? Just announced! 🙂
Prentis: The Agents That Learn to Use a Computer
Most of the labs in this series are chasing intelligence in the abstract — better reasoning, better world models, better language. Prentis is chasing something narrower and, arguably, more immediately commercial: agents that can operate a computer the way an office worker does.
Founded in April by serial entrepreneur Ritankar Das alongside LinkedIn co-founder Reid Hoffman and Zynga co-founder Mark Pincus, Prentis is now reportedly in talks to raise $100 million at a $1 billion valuation.
The premise is straightforward. A huge amount of white-collar work is not creative or strategic — it is routine navigation through documents and internal systems: processing insurance claims, chasing paperwork for customs duty refunds, moving data between forms. Prentis is training models to learn those workflows directly, with the goal of building agents that control a computer end-to-end rather than answering questions about one.
The company claims real commercial traction ahead of most research-stage labs in this series. It has signed contracts worth up to $50 million with customers including a healthcare management service organization, a manufacturer, and clothing and goods manufacturers. Investor materials reportedly project a $75 million annualized run rate by Q3 of this year — though Prentis’s own pitch deck is careful to caveat that figure as estimated value based on a 20%-of-savings fee structure, not recognized revenue, and “performance-dependent and subject to final execution.”
On the technical side, Prentis says its Hive-32B model beats GPT-5.4 and Claude Opus 4.6 on two computer-use benchmarks — WindowsAgentArena (end-to-end task completion in real Windows applications) and ScreenSpot-v2 (locating the correct on-screen control). The pitch is not “we built a smarter model” but “we built a smaller, cheaper one that’s good enough”: Prentis claims roughly 10x lower cost per task than frontier APIs, which matters enormously if the goal is deploying agents across thousands of repetitive back-office workflows rather than a handful of high-value queries.
None of this is independently verified. The benchmark claims come from Prentis itself, and the revenue figures come from a pitch deck built to raise money.
But the strategic bet is worth naming clearly, because it’s a different bet than almost every other lab in this series is making. Frontier labs are racing to build the most capable general model. Prentis is betting that for a huge category of real economic work, the winning model isn’t the smartest one — it’s the cheapest one that’s reliable enough to run unsupervised, all day, on someone else’s screen.
If that thesis is right, “computer use” becomes its own competitive category, separate from the reasoning and multimodal races most of this series has covered — and one where incumbents’ general-purpose advantage matters less than specialization and cost…
Originally published on my main website HardTech Reads: The AI & Robotics Revolution
🤝 Passionate about HardTech? Have comments, want to share some thoughts or ask about sponsorships? Let’s connect!
👉 Want to feature your company in HardTech Reads? Apply Here
👉 Want to schedule a meeting directly, check my schedule.
No Comments