Weekly AI x Robotics digest: open models, AI capital, Unitree IPO
Thread1,841 words
First weekly digest on AI x Robotics: open-weight releases (GLM-5.3, DeepSeek V4 Pro, Qwen3.8-Max), AI funding and revenue, and the Unitree IPO. Argues post-training is the new moat and Hyperliquid perps priced Unitree better than the Shanghai open.
Key claims
Anthropic annualized revenue $65B as of July vs $47B in May
Unitree opened at 1,100 yuan, closed day one at 845, fell about 17% on day two
GLM-5.3 weights delayed to about Aug 28 over emergent cyber capability
AI x Robotics is the super-cycle that you've been waiting for.
But it's hard to keep track of everything that is happening
That's why I'm gonna start a weekly series to cover the key news in the space here on X.
First edition ↓ t.co/qVnQ6V2eH4
1 ) AI Models
1.1 ) GLM-5.3 by @Zai_org
Same 743B base as GLM-5.2, with all gains coming from post-training. Terminal-Bench 3.0 from 4.6 to 28.3, DeepSWE from 46.2 to 66.9, CyberGym 84.5. API went live this week at an unchanged $1.40/$4.40 per M tokens. @ArtificialAnlys has it at 60, tied with Kimi K3 for best open-class model. Post-training could be the new moat. Also made a post about this here: x.com/zerokn0wledge_/status/2088426954441601116.
1.2 ) GLM-5.3's emergent cyber problem
Argubly the bigger story than the benchmarks above. Offensive skills scaled a bit faster than @Zai_org planned lol (from bug finding to full exploitation chains with 2,436 real vulns found across 269 projects, 1,097 medium-to-high, hitting Linux, WebKit, FreeBSD). Weights hence delayed to ~28th August for safety hardening. It's the first time a Chinese lab gates a release over an emergent capability. Open-weight models could turn into an accelerator for the cyber threat landscape.
1.3 ) DeepSeek V4 Pro fully live (mid-August)
New flagship model of @deepseek_ai is out of preview. 1.6T MoE, 49B active, 1M context, with new peak/off-peak pricing since Aug 16. Better agents and a price hike in the same release. A clear sign that the "cheap labs" won't necessarily stay cheap forever.
1.4 ) Meta reopens the weights (Aug 10)
@finkd's "The Future Is for Everyone" essay came with Muse Glimmer: 30B, Apache 2.0, small enough to run on a consumer GPU. First open @Meta weights since Llama 4 in spring 2025. The flagship Muse Spark 1.2 is supposed to follow "in the coming weeks".
1.5 ) Qwen3.8-Max: biggest open-weight model ever (Aug 3)
2.4T total, 95B active, 1M context. First open-weight Max-class Qwen (by @Alibaba_Qwen) ever. 86.6 on Terminal-Bench 2.1, still trails the frontier on the hardest coding evals but is closing the gap rapidly. The 27B version on the other hand runs well on local infra, and is def one of the better choices when hardware is a constraint and good to see China bring super-intelligence to our desks through open models. Wrote about this here: x.com/zerokn0wledge_/status/2084187995754639848.
1.6 ) MiniMax H3: open weights with a geofence
One 33.1B model that handles text, image, video and audio in the same context and generates 15 second clips with sound. Weights are on Hugging Face and the ecosystem moved fast to support it (incl ComfyUI, CapCut, Leonardo, AMD). The catch sits in the license tho as the open version tops out at 768p, the 2K upscaler is API only and usage is restricted for local use in many jurisdictions.
2 ) Capital and Business Performance
2.1 ) Anthropic revenue metrics up ahead of approaching IPO
In its May Series H @AnthropicAI raised $65B at $965B post-money valuation. Now it reached $65B in annualized revenue as of July per @Bloomberg, up from $47B in May and ~$9B end of 2025. Q2 is already above $11.5B with a first positive adjusted operating profit. S-1 for IPO is filed, and listing is reportedly possible around late September. You can already trade pre-IPO perps on Hyperliquid btw: x.com/zerokn0wledge_/status/2090500183893442733.
2.2 ) OpenAI revenue back above $40B
OpenAI is now making money at a pace of over $40B per year as per Bloomberg reporting. That is double where they ended 2025, after growth stalled around $25B in spring. Actual Q2 revenue: $6.7B. Impressive, but (as outlined above) still less than Anthropic, who sits at a $65B yearly pace.
2.3 ) Nvidia x SB Energy x OpenAI: The Ohio megadeal is signed (Aug 17)
The deal behind @OpenAI's giant Ohio data center campus finally closed, confirmed in SEC filings. The structure: SoftBank's SB Energy builds and owns the site, OpenAI rents it for 20 years, and every chip inside is @Nvidia. To make it happen, Nvidia invests $1.5B directly and guarantees up to $105B of the financing. Weeks ago the number being discussed was $250B, so Nvidia cut its risk by more than half, and the stock rose on the smaller commitment. Nvidia also says the OpenAI relationship could be worth around $600B in chip sales through 2030.
2.4 ) Stripe buys OpenRouter (Aug 19)
Confirmed by both @stripe and @OpenRouter at an undisclosed price, but with reported prive ranging from $7.5B to $8B+. OpenRouter has seen surging adoption recently with 10T+ tokens routed daily across 400+ models. Means payments company that also builds stablecoin payment rails with @tempo now owns the token metering layer too. Interesting basis for agentic payment and model routing infrastructure.
2.5 ) Etched doubles to $21B in 26 days (Aug 18)
The AI chip startup @Etched raised $700M led by @JaneStreetGroup at a $21B valuation. It was worth $10.3B in July and $5B in December n(crazy how fast valuations currently rise in AI industry). What makes this round different tho is that Jane Street is not just the lead investor, they are also the first customer. They got a rack of Etched chips in July, ran real trading workloads on it, and then went on to lead the round.
2.6 ) Databricks raises $5B at a $190B valuation (Aug 13)
Led by @coatuemgmt, @databricks completed its second $5B raise this year, up from a $134B valuation in February. The business behind it is real: about $7B in yearly revenue, up over 80%, and they generate more cash than they burn already. At current valuation of $190B, investors are paying 27x revenue for the picks and shovels of enterprise AI.
2.7 ) Google buys a dead airline's data for $10M (Aug 14)
Spirit Airlines went bankrupt, so its data went up for auction. @Google won with $10M, including 100 million internal emails, 500 million Teams messages, billions of pricing and booking records, and files on 175K employees going back to 1986. Interesting precedent. When a company dies, its data is now an asset worth fighting over.
2.8 ) Gemini app crosses 1B monthly users (Aug 11)
One billion people use the @GeminiApp every month. Fastest growing product Google has ever shipped. Evidence to the fact that AI assistants stopped being an experiment, and are now real mass-market infrastructure.
3 ) Robotics Season
3.1 ) Unitree: +460%, then gravity (Aug 19 to 20)
The first humanoid robot maker ever listed in mainland China went public and the open was insane. $UNITREE (688836) started trading at 1,100 yuan, up 629% from the 150.80 IPO price, and closed day one at 845, worth about $50B. All of this while the rest of the Shanghai tech index dropped 7.2% the same day. On day two tho, the stock fell about 17% to around 700 yuan, still equalling a roughly $40B company valuation. Context for the frenzy: only 7.4% of shares were actually tradable after lockups, and the IPO was heavily oversubscribed due to enormous retail demand (more here: x.com/zerokn0wledge_/status/2088992749571674162). At the day one high the stock traded at roughly 857 times next year's expected earnings btw (while the Shanghai market average is about 130). Clear evidence to the massive @UnitreeRobotics hype.
3.2 ) Hyperliquid perps called the Unitree print
Before the stock ever traded on SSE, pre-IPO perps on @HyperliquidX priced Unitree at $90-$100 per share, implying a ~$38B-$40B company. That looked wildly aggressive many argued, as that was already 4-5x above the IPO price. Then the stock listed, spiked, and by day two traded right back to the perp levels. Crazy to think that a crypto derivatives market priced a Shanghai IPO more accurately than the Shanghai opening auction did, and proof that Hyperliquid is the leading pre-IPO price discovery venue and IPO oracle rn.
3.3 ) Unitree fundamentals: real revenue, real policy risk
This is not a story stock with no business. 2025 revenue came in at CNY 1.70B, up 335%, with CNY 278M in net profit. Actually profitable, which is rare in this sector. Around 18,000 humanoid robots delivered to date, over 5,500 last year alone. Tencent, Alibaba and DeepSeek are all on the cap table. Now the problem: 44% of revenue comes from overseas, and the US just slammed that door. The FCC banned new foreign humanoid imports in late July and the Pentagon has Unitree on its list of Chinese military companies. So the market is paying a record multiple for a company whose biggest growth market is being closed by regulators. For deeper coverage of the fundamentals, check this report by @RoboStrategy: x.com/RoboStrategy/status/2087561451468681234.
3.4 ) "Superman": the IPO week flex (Aug 17)
Two days before listing, Unitree unveiled a new robot claiming a 2 meter standing jump and a top speed of 12.66 m/s. That is faster than Usain Bolt at his peak. Fun marketing stunt to capture attention around IPO.
3.5 ) World Robot Conference: deployment season (Beijing, Aug 19 to 23)
Beijing's biggest robot show opened the same day Unitree listed, 300+ companies, 150+ new products. The vibe shifted this year: less dancing robots, more robots actually working. Parcel sorting, factory lines, elder care. The numbers behind it: about 19,000 humanoids shipped in the first half of 2026, and 97% of them came from China. One honest caveat: analysts say a big chunk of those units are collecting training data rather than doing paid work. Meanwhile the IPO line behind Unitree keeps growing, AgiBot is filing in Hong Kong and Lumos and Chery's robot arm are considering listings (covered the IPO pipeline here: x.com/zerokn0wledge_/status/2090010902523187653).
3.6 ) Figure: the US counterweight
While China lists robot companies publicly, America's best-funded one stays private. @Figure_robot built its 1,000th Figure 03 robot in July, moved its BMW deployment from a body shop pilot to actual logistics work, and signed retail group Catalyst Brands. That is the most documented commercial track record in Western robotics. But it sits inside a private company marked at $39B where nobody outside can check the numbers.
3.7) Dyna-2: human video as the robotics scaling law (Aug 10)
Probably the most important robotics research piece of the month. Dyna trained a robot model on over 1 million hours of video shot from the human point of view (e.g. people cooking, folding, assembling) with zero robot data in the initial training. The claim: robot skill improves predictably as you feed it more human video, the same scaling logic that built LLMs. Their numbers show task success jumping from 20% to 90% purely from more video, and a robot learning to open a bottle cap after just 13 minutes of demonstration. Could be a big step given the amount of video data the internet offers vs. how scarce robotics training data is.
3.8) Everything else that's happening
Serve Robotics is putting delivery robots on sidewalks for Grubhub. @PonyAI_tech is lining up thousands of robotaxis outside China. North American factories ordered $622M worth of robots in Q2, up 21% from last year, and more and more crypto x robotics projects are emerging. Covered the latter here: x.com/zerokn0wledge_/status/2088387258940268823.
4 images on X