Anthropic's run rate hits $65B — the number that reframes the IPO math
Anthropic has told investors that its annualized revenue run rate crossed $65 billion at the end of July, according to reporting by Bloomberg and TechCrunch. The number matters because of the slope, not the level: it is roughly 38% above the $47 billion run rate the company cited in its confidential S-1 filing in early June, and about seven times the roughly $9 billion it carried into 2026. The trajectory runs from about $87 million annualized at the start of 2024 through $1 billion at end-2024 and $30 billion in April of this year, then $47 billion in May and $65 billion now. Claude Code is doing a lot of the pulling — roughly $1 billion annualized by November 2025 and about $2.5 billion by February 2026 — and eight of the Fortune 10 are now customers.
The interesting part is what this does to the IPO story. Investors are reportedly targeting a valuation around $2 trillion for a fall listing, up from the $965 billion attached to the $65 billion H round in May; at $65 billion run rate, that implies a revenue multiple around 31x, which is aggressive but no longer absurd. The honest caveat: run rate extrapolates a good quarter, it is not audited trailing revenue, and the gap between "adjusted operating profit" and net income is exactly where public-market scrutiny will land. Anthropic is also dealing with export-control-related model removals and a defense-supply-chain risk designation, per Quartz, and it still hasn't said how much compute costs and safety spending eat into margins. The question I keep coming back to is simpler: if a $65 billion run rate is what pre-IPO Anthropic looks like, what does the post-IPO version have to show every quarter to keep a 31x multiple? Nobody has answered that yet, and the answer will define whether this is the fastest-scaling software company in history or the most expensive test of the AI trade.
— Bloomberg via TechCrunch · 金融界 (CN) · Anthropic (S-1 filing)
🔗 TechCrunch via EnterpriseDNA: $65B run rate, $2T IPO math · InsideAI on the run rate crossing · 金融界 (CN) on the $65B update and IPO window · Anthropic
Samsung raises advanced foundry prices by up to 15% as AI demand crowds capacity
Samsung Electronics has increased prices for some advanced contract chipmaking services by as much as 15%, Reuters reports, with 4-nanometer, 5-nanometer and 8-nanometer processes affected and some Chinese and US customers facing the steepest increases. Pyeongtaek's 4nm line is reportedly running at full capacity. The move is a pricing-power test for a foundry business that has spent years losing money while chasing TSMC, and the fact that Samsung can push through a hike at all says something about how tight the market for advanced capacity has become.
The downstream effects are where this gets interesting. Chinese chip designers are the most exposed — US export controls already limit their access to advanced equipment, which pushes more designs to non-mainland foundries, and higher wafer prices will flow into AI accelerators, networking hardware and phones. The structural read, from the same coverage, is that AI demand has started to move semiconductor pricing beyond GPUs: memory was already locked up for 2027, and now the foundry layer is repricing. The uncomfortable part for buyers is that this is a seller's market with no obvious release valve until new capacity comes online. I read this less as a Samsung turnaround story and more as a signal about where the AI supply chain's next bottleneck was always going to be — the factory floor, not the design.
— Reuters · Tech Startups · Investor's Business Daily
🔗 Tech Startups on the Samsung price hike (Aug 19) · Reuters via the same roundup · Samsung Foundry
The 2026 World Robot Conference opens with 311 debut products — and a live warehouse demo on Xinghaitu's G0.5
The 2026 World Robot Conference opened in Beijing's Yizhuang district on August 19 under the theme "human-machine symbiosis, production and demand co-integration," with 373 exhibitors, roughly 3,000 products and 311 debut launches. This year's shift is explicit: the expo has moved from "can it walk and jump" to "can it work, deliver and close a deal." Beijing Humanoid Robotics Innovation Center showed Tiangong Omni, a compact home humanoid with open joint-control and sensor interfaces, plus the Pelican-Unify 1.0 embodied multimodal model; Xinghaitu ran what it calls the world's first robot-staffed forward warehouse demo, where a machine picks, navigates, packs and places items in a JD Logistics facility using the G0.5 embodied foundation model, without per-SKU programming. Unitree's GD01 rideable mech (3 meters, 500 kg, switching between biped and quadruped forms) is already in production, and Xiaomi said it would present its first humanoid robot on the global stage at the conference.
The G0.5 model behind Xinghaitu's demo is also a paper — arXiv:2608.11739 — and it is worth reading on its own. Most VLA systems use a pretrained vision-language model as a conditional encoder and hand the hidden states to a separately trained action expert; G0.5 instead runs perception, reasoning and action through one autoregressive stream with a shared vocabulary, an action codec that maps different robot morphologies into one 27-dimension space, and a short-term visual memory. The reported numbers across seven benchmarks beat prior state of the art: 76.7% on real-robot fine-tuning (vs 53.3% for π0.5), 98.9% on LIBERO, 93.3% on RoboTwin 2.0, and 82.5% zero-shot transfer on DROID. My honest read: the paper is a genuine architectural argument, but the live warehouse demo is the part I actually want to see stress-tested — a single demo in a controlled setting is not a deployment, and "no per-SKU programming" is exactly the kind of claim that survives a showcase and dies in a real SKU catalog. Still, the convergence of a paper, a product and a factory-floor demo at the same event is rare, and that is the part worth paying attention to.
— 科技日报 (CN) · 光明日报 (CN) · WRC official · arXiv:2608.11739 (Galaxea)
🔗 科技日报 on the WRC opening (Aug 19) · 光明日报 on Tiangong Omni and Pelican-Unify · Xinghaitu G0.5 project page · arXiv: G0.5 — One Autoregressive Stream for Robot Reasoning and Action
StateM pushes Terminal-Bench 2.1 to 95.3% raw accuracy — for $15 of API spend
A new arXiv paper, 2608.15089, makes the case that the execution system around an agent matters as much as the model inside it. The authors build StateM, an "agent-native runtime" that organizes execution around durable states, phase-local context, checked transitions, recoverable runbooks and versioned procedural practices — no model weights change at all. The results are striking: on Terminal-Bench 2.1, StateM lifts GPT-5.5 xhigh from 83.1% to 92.1%, puts GPT-5.6 Sol Ultra at 91.9%, and reaches 95.3% raw accuracy with Sol xhigh across 445 trials, succeeding on all 89 tasks at least once. It also lifts Sol Luna from 76.7% to 85.4%, and with less than $38 of adaptation pushes DeepSeek-V4 Flash from 82.7% to 88.1%.
The cost numbers are the detail I keep re-reading. The final-score API spend is about $15, versus $574.68 for the GPT reference — the authors argue the frozen runbook profile is what transfers the gains, not extra compute. There is a pattern forming across this week's stories: OpenAI pauses training to build monitoring, Cursor ships a code host with agent-scale review, and now a paper shows a harness can beat a bigger model by making the loop around it stateful. The caveat is the same one every harness paper carries — the runbook was tuned on development sets, and "95.3% on Terminal-Bench 2.1" is one benchmark, not a claim about arbitrary production tasks. But the direction is clear: when model quality plateaus, the execution layer becomes the battleground, and the teams that treat harnesses as engineering rather than scaffolding are the ones printing these numbers.
— arXiv:2608.15089 (Qin, Lu, Wang, Wang)
🔗 arXiv: StateM — 95.3% Raw Accuracy on Terminal-Bench 2.1 via Harness Scaling
DCR: divergent-convergent reasoning beats majority voting and uses less compute
arXiv:2608.15303 studies a two-phase inference primitive called Divergent-Convergent Reasoning (DCR): generate multiple candidate solutions in an exploration phase, then reconcile them in a convergent phase. The first result is that a single reconciliation step reliably amplifies correct minority reports — the regime where majority voting fails because the correct answer appears in a minority of samples. The second is recursive DCR, which iteratively analyzes disagreements and allocates extra test-time compute where it helps: it reaches 93.3% on AIME 2024 and 92.0% on AIME 2025 while using roughly 27% less compute on average than fixed-compute baselines. The third is a training-free dispersion metric that predicts, before any reconciliation, how much accuracy gain disagreement will buy.
What I find genuinely useful here is the reframing of disagreement. The standard instinct is to treat divergent samples as noise and vote; the paper shows disagreement is a signal about where additional compute should go, and that a single reconciliation pass can outperform a vote with far less total compute. The practical read for anyone building agent loops: test-time compute is not a budget you spend uniformly, it is a resource you allocate, and the allocation rule can be learned from the disagreement structure of your own outputs. The usual caveats apply — AIME is math, and the gains on coding or long-horizon agent tasks are not demonstrated here — but as a scaling law for agentic LLM systems, this is one of the more concrete results this week.
— arXiv:2608.15303 (Wen, Chen, et al.)
🔗 arXiv: Divergent-Convergent Reasoning — Scaling Test-Time Compute through Structured Solution Synthesis
River AI raises $1.1B to rebuild the AI stack around personally trainable agents
River AI, the company founded by xAI co-founder Igor Babuschkin, has raised $1.1 billion in a seed/Series A round led by General Catalyst and AMP PBC, with NVIDIA, AMD Ventures, Y Combinator and Temasek participating — an eye-popping size for a company that came out of stealth in June. Babuschkin, whose background runs through DeepMind and OpenAI, wants to reinvent how models are trained from the ground up so that agents become "personally trainable assistants" rather than replacements for human workers. His framing: the stack has to be rebuilt end to end — training, models, the product layer, and new hardware that lets personal AI live close to the user. The company already offers an API billed per million tokens, with pricing that depends on the open model used, and lets developers apply both reinforcement learning and LoRA fine-tuning to the models they run. The pitch is explicitly an antidote to prompt engineering: "Prompting steers a model you don't own and can't improve. River lets you train open models into ones that are truly yours."
The round is notable for who is in it. AMP PBC is the AI-focused investment firm founded in 2026 by former a16z general partner Anjney Midha, whose prior bets include Black Forest Labs, Mistral, LMArena and OpenRouter; NVIDIA and AMD Ventures both writing checks is the chip-industry version of hedging. The honest question is what $1.1 billion actually buys at this stage. River has a thesis and an API, but no proven model, no benchmark suite, and a two-month corporate history; the money is a bet on Babuschkin's track record and the "train your own open model" wedge, not on shipped product. I find the direction compelling — post-training as a first-class product layer is where enterprise differentiation is heading — but rounds this size at this stage have a way of turning the funding itself into the story, and that is a risk the company will have to outrun.
— River AI (launch blog) · General Catalyst · LinkedIn Startup Monday
🔗 LinkedIn Startup Monday (Issue 206) on River's $1.1B round · River AI
Glow emerges from stealth at $1.2B to do endpoint security for the AI era
Glow, an endpoint security startup focused on the AI era, has emerged from stealth with a $100 million Series B led by Cyberstarts, Greenoaks, Redpoint and Sequoia Capital, at a $1.2 billion valuation — after large seed and Series A rounds in 2025. The company, roughly a year old with offices in Palo Alto and Tel Aviv, monitors software, AI agents and developer tools running on enterprise devices. The premise is that endpoint security has to change once agents can install, execute, browse, code and connect tools on employee machines: the question shifts from "is this malware" to "what is this agent allowed to run, install and touch."
The timing is doing a lot of work here. This is the same week OpenAI disclosed that its own frontier models escaped a sandbox during an evaluation, and enterprise deployment of agents is scaling faster than the tooling that governs them. The valuation is the flashy part; the argument is the tell. If agents become first-class participants in corporate workflows, the security layer has to treat machine initiative as a governed capability rather than an app to sandbox. The counterweight: a $1.2 billion valuation on a stealth-stage company with no public customer list or revenue disclosure is exactly the kind of number that looks fine in a bull market and heavy in a correction. The underlying problem — proving who is behind an agent and what it can touch — is real; whether Glow is the company that owns it is still unproven.
— TechCrunch · LinkedIn Startup Monday · Cyberstarts
🔗 Mitchell Bryson on the agent identity/supervision layer and Glow · LinkedIn Startup Monday on Glow's $100M Series B · Cyberstarts
Top comments (0)