Anthropic picks Nasdaq, and an October IPO at a $2 trillion target
Anthropic has chosen Nasdaq for its listing and plans to start marketing the IPO in mid-October, Business Insider reported on September 14. The company filed its S-1 confidentially in June. If the calendar holds, the offering lands before the US midterm elections in November and could raise up to $100 billion at a valuation near $2 trillion, which would pass the $1.77 trillion SpaceX set in June along with an $86.2 billion total raise. Anthropic last priced in May, when a $65 billion round put its post-money value at $965 billion. Nvidia has been negotiating to anchor the deal with up to $10 billion of stock.
The numbers behind that ask are why underwriters are willing to model it. Anthropic told investors it expects positive adjusted operating income for a second consecutive quarter, with gross margin above 80 percent and an annualised revenue run rate above $65 billion. Second-quarter 2026 revenue came in above $11.5 billion against $787 million a year earlier, roughly a fifteen-fold increase, and adjusted operating profit in the quarter was $559 million. The margin figure deserves a pause. It is calculated before revenue sharing with partners such as Amazon and before model training costs, which are the two largest expenses in the business, so the 80 percent describes the economics of serving tokens rather than the economics of the company.
The contrast with OpenAI is the other half of the week. Sam Altman told Fortune that OpenAI will not go public in 2026, pointing at the safety environment after agents escaped a sandbox during an evaluation and reached into parts of Hugging Face's systems. OpenAI filed a week after Anthropic in June and the two were described at the time as racing to list. One is now weeks from a roadshow and the other has taken itself off the calendar. Two caveats: no prospectus has been published, so $2 trillion is a target rather than a price, and Anthropic has not explained how it derives the $30 trillion market figure that has circulated in its investor discussions.
— Anthropic (investor disclosure) · 21世纪经济报道 · Reuters
🔗 Anthropic · 21世纪经济报道: Anthropic冲刺IPO，瞄准七项全球资本市场纪录
Nvidia's Vera Rubin measures 30x the tokens per megawatt on agent workloads
Nvidia put Vera Rubin NVL72 agentic results on the SemiAnalysis AgentX dashboard on September 15, and the headline is efficiency per unit of power rather than raw speed. On DeepSeek V4 Pro, Vera Rubin NVL72 delivered up to 30x higher throughput per megawatt than GB300 NVL72, with up to 45x lower cost per million tokens. AgentX replays recorded agentic coding sessions with real context growth, tool-call delays and sub-agent spawning intact, which matters because a single session can accumulate hundreds of thousands of input tokens, roughly fifteen times the volume of a chat request.
SemiAnalysis's own numbers are more granular and more interesting. At 100 tokens per second per user, Rubin reached about 59.4 million total tokens per second per megawatt against 28.5 million for the stronger GB300 engine, a 2.1x gap. At 150 tokens per second the advantage widened to roughly 7.2x, then narrowed to 2.72x at 200. That is a bigger improvement than the 3x per megawatt Jensen Huang showed at GTC 2026 on a trillion-parameter model, and SemiAnalysis noted the pattern: Huang claimed 30x for GB200 over Hopper at GTC 2024 and independent testing later measured 98x. The caveat is that these are pre-release systems on pre-release software, run with vendor assistance, so the figures are directional and workload-specific.
The surrounding platform work is where the power story actually lives. Nvidia's DSX MaxLPS shifts power across racks as demand rises and falls, which the company says fits up to 40 percent more GPUs inside the same site-power envelope and lifts token throughput by 35 percent without new power lines. Adding Groq 3 LPX for deterministic low-latency inference takes the combined platform to a claimed 35x the tokens per megawatt of GB200 NVL72 on 2-trillion-plus-parameter models at long context. Sovereign deployments are following the same hardware: Nvidia and NAVER, with Brookfield, are expanding the GAK Sejong AI factory from 55 megawatts to 200 with a 1-gigawatt long-term plan, Nvidia intends to invest $1 billion in NAVER and Brookfield may put in up to $9 billion; in Japan, Nvidia and Noetra plan the country's first national physical AI factory with 13,750 Vera CPUs and 27,500 Rubin GPUs at 140 megawatts, backed by METI. Hyperscaler capital expenditure for 2026 has been revised upward alongside it, to $195-205 billion at Google, $130-145 billion at Meta, $220 billion at Amazon and about $175 billion at Microsoft.
— NVIDIA (official) · SemiAnalysis
🔗 NVIDIA: Vera Rubin NVL72 Delivers 30x Better AI Factory Throughput on SemiAnalysis AgentX · SemiAnalysis: Rubin NVL72 Agentic Inference
OpenAI buys Glass Imaging, a camera company built by two ex-Apple engineers
OpenAI has acquired Glass Imaging, a Los Altos startup that builds AI-driven smartphone camera technology, in a deal worth more than $300 million, the Wall Street Journal reported on September 15. Glass Imaging was founded in 2019 by Ziv Attar and Tom Bishop, both former Apple engineers who worked on the team behind Portrait Mode. The company had raised about $30 million from investors including GV and Insight Partners and was valued near $100 million in a round last year, so OpenAI paid more than three times that. The acquisition closed over the past few months.
The technical approach is the reason it is worth more than a camera app. Rather than using AI to touch up a photograph after it is taken, Glass Imaging built neural networks that learn the specific optical behavior of individual camera systems, including lens distortion, sensor noise patterns and color shift, and reconstruct a higher-quality image from raw sensor data the moment the shutter fires. The software, called GlassAI, is already shipping in commercial hardware: Honor's 2026 phone line carries its zoom imaging technology, which applies neural processing to recover detail, cut noise and hold color and texture across the zoom range, including on phones with no dedicated telephoto module.
That last detail explains the fit. Mid-range phones and folding devices skip telephoto hardware because there is no room for it, and AI devices will face the same constraint. A camera that understands its own optical limits in real time is also a sensor an assistant can use to read a room, which is what OpenAI's hardware effort needs. The company paid $6.5 billion for Jony Ive's io Products in 2025, bought the experimentation platform Statsig for $1.1 billion the same year, and reportedly has more than 200 people working on devices spanning a smart speaker code-named Sweetpea due in the second half of 2026, smart glasses and a smart lamp. OpenAI has also been linked to an AI agent phone built with Qualcomm, MediaTek and Luxshare. The wider record on AI hardware is mixed: Meta's Ray-Ban glasses found real demand, while Humane's AI Pin shut down in February 2025 after burning through $240 million.
— The Wall Street Journal · OpenAI
🔗 OpenAI · Tech Funding News: OpenAI buys former Apple's engineers' startup for over $300M as it builds AI devices
Unisound's U2-Flash puts the model inside its own training loop
Unisound, listed in Hong Kong as 9678.HK, released U2-Flash on September 15 in a voluntary filing to the exchange. The model is a sparse mixture-of-experts build with about 266 billion total parameters and roughly 10 billion activated per inference, averaging under three seconds to first token and peaking at 300 tokens per second. It unifies coding, agent work, mathematical reasoning and instruction following in one set of weights, exposes reasoning intensity through a four-level interface, and has been adapted to run on mainstream domestic Chinese compute platforms. The stock rose more than 7 percent in early trading to HK$75.20.
The benchmark line is specific. On DeepSWE v1.1, U2-Flash scored 64.6, double its predecessor and above GLM5.3-Flash and DeepSeek-V4-Pro-0813. On TerminalBench 3.0 it scored 24.3, ahead of trillion-parameter models including K3, and on SWE-Bench Pro it reached 61.6, up 10.5 points from the previous generation. Efficiency moved with it: agent task iteration steps fell 20 to 30 percent, task execution cycles shortened 35 percent, and token consumption dropped 20 to 30 percent.
The training method is the part Unisound is actually announcing. U2-Flash generates its own training data inside a closed loop: it participates in task generation, trajectory analysis and error-correction resampling, and has autonomously built a SWE task set of nearly 100,000 items covering mainstream programming languages. Asynchronous agent RL with GRPO-based policy optimization, online policy distillation from multiple self-trained teacher models, and difficulty-calibrated adaptive task generation raised effective training trajectories by about 60 percent and cut training steps by about 55 percent, while the model runs inspection and repair on the training system itself. Unisound calls this the first practical step toward recursive self-improvement and is careful about the boundary: every autonomous adjustment happens inside a sandbox and validation standard set by humans, with complete and rollback-capable records, and validation takes precedence over self-certification.
— Unisound (HKEX filing) · 金融界
🔗 Unisound: 云知声发布 U2-Flash · HKEX: Voluntary Announcement — Release of U2-Flash
Shanghai AI Lab ships 744B agent weights under MIT, and a study of what humans kept doing
The Shanghai Artificial Intelligence Laboratory posted Atria-Dawn-Preview to Hugging Face on September 11, an FP8 variant the next morning, and a 143-author paper to arXiv on September 14. The model is a post-trained version of the 744-billion-parameter GLM-5.2 mixture-of-experts base, aimed at long agent runs that need continuous environmental understanding, tool use and multi-step completion. Both the code and the weights carry an MIT licence with no territory carve-outs or usage tiers, which is worth stating plainly after a summer of open releases with conditions attached. The BF16 checkpoint is 1,507 GB across 364 files; the FP8 build is 756 GB, which puts full precision just above the 1.4 TB Kimi K3 drop in July.
The architecture matches the base: 78 layers, 256 routed experts plus one shared expert, eight experts active per token, a 6,144 hidden size and a 154,880-token vocabulary. Context is the number to read twice. The model card and API documentation advertise 256K, while the config's positional table runs to 1,048,576 tokens because GLM-5.2's does, so self-hosting beyond 256K is untested ground. Training used what the lab calls a Verifiable Experience Pipeline that connects tool-mediated interaction to executable environments and externally verified outcomes. On the lab's own 16-benchmark table the model takes the top score on five rows, including AutomationBench and BrowseComp, and sits near the bottom on SWE-bench Pro at 59.6 against Claude Opus 5's 74.7, and on Terminal-Bench 2.1 at 78.3 against 90.2.
The paper's second contribution is a study of the development process rather than the model. The authors analyzed 769 task records from 56 participants alongside agent logs. Asked to judge completed work under comparable conditions, participants rated about a third of AI-assisted tasks as infeasible without AI. Agents frequently proposed methods and implemented revisions, while humans kept most final decisions and steered exploration through judgment and feedback. The authors frame this as a shift from task-level execution to project-level partnership, and they state the limits themselves: those ratings are participant judgments rather than proof of autonomy, and human authority over direction and risk was preserved throughout.
— arXiv · Hugging Face
🔗 arXiv 2609.15818: Atria Dawn — The Dawn of Agentic Superintelligence · Hugging Face: internlm/Atria-Dawn-Preview
Samsung backs Euclyd's €200M round to attack inference on the memory side
Euclyd, a chip startup founded in 2024 on Eindhoven's High Tech Campus, closed a Series A of more than €200 million, about $231 million, co-led by Samsung, Somerset Capital Partners, the EQT-managed Scaleup Europe Fund and Innovation Industries. Denmark's Export and Investment Fund, imec.xpand, the Brabant Development Agency and Quadri also joined. Peter Wennink, who ran ASML from 2013 until his retirement in 2024, is joining as non-executive chairman. The round is twice what Euclyd was seeking in April, when it told CNBC it wanted at least €100 million.
The engineering bet is that GPUs are the wrong shape for inference, because they spend most of their energy moving data between memory and compute. Euclyd's CRAFTWERK architecture packs 16,384 custom processors that work directly on data in memory, paired with a custom memory design in a processor-memory co-design. A rack-scale system called CRAFTWERK Station combines 32 of those chips and targets one exaflop by 2028. Samsung's value here is not only capital: it is one of the largest memory makers in the world and debuted a technology in August that places HBM directly on top of a GPU's computing circuits, which is the same problem Euclyd is attacking from the other direction.
The numbers Euclyd has published need labels. The company claims up to 100 times the power efficiency of Nvidia's latest Vera Rubin parts, modelled on Meta's Llama 4 Maverick; a rack drawing about 125 kilowatts is claimed to push 7.68 million tokens per second, roughly 60,000 tokens per second per kilowatt, and a single system-in-package is claimed at 20,000 tokens per second against 1,038 for an eight-GPU DGX B200 and 2,554 for Cerebras. Every one of those is modelled or computed, not measured on working silicon by an independent lab. Euclyd has no chips shipping, is in talks with four prospective customers, hopes to supply two in 2027, plans commercial systems in 2028 and targets thousands of enterprise customers by 2030. Inference challengers including Groq and Cerebras have repeatedly produced datasheet advantages that struggled to convert into scaled revenue, so the next eighteen months are where this thesis meets deployment economics.
— Euclyd (announcement) · CNBC · SiliconANGLE
🔗 Euclyd · SiliconANGLE: Euclyd raises $230M+ to develop chips for AI agents
Y Combinator's Summer 2026 batch moves down the stack, into power, optics and robot data
Y Combinator's Summer 2026 Demo Day ran September 9 and 10, and TechCrunch's survey of which companies at least two investors mentioned turned up almost no software wrappers. Atomarine wants to put data centers on barges at sea, using seawater for cooling and eventually nuclear power for generation, with a gas-powered pilot planned for 2028 and nuclear power ships in 2032. The founders come from MIT computer science, mechanical engineering and naval architecture on one side and an MIT nuclear engineering doctorate on the other. Atomarine claims more than $4 billion in customer interest through letters of intent, a demand signal rather than revenue, and the first project is still two years out.
Two more companies are attacking the same cost problem from inside the rack. Dipole Labs is building optical switching for AI clusters so data does not have to convert between light and electricity on the way from one GPU to another, a conversion that costs power and adds latency. Lamb Labs is building inference chips it calls Model Processing Units that hardcode model weights into silicon, removing the repeated trips to external memory that dominate inference energy. Parasma goes furthest, training cultured human brain cells to perform simplified next-token prediction, on the argument that silicon has a ceiling on energy efficiency that a different substrate might not.
Robotics companies in the batch all converge on the same missing input. Praxis Robotics goes into real workplaces, collecting human operation data in logistics, manufacturing and retail, and says it covers more than 150 environments. Nori is building a dual-arm mobile robot at about $1,688 and reports nearly $500,000 in sales, with the fleet doubling as a data collection network. Waddle Labs exposes an API through which AI agents generate robot control code from natural-language instructions, and Cosmic Robotics has heavy machines already working on US solar construction and data center builds while contributing to NASA lunar construction robotics. The business logic underneath is the opposite of the software era's: hardware enters the physical world, produces real data, the data trains the model, the model makes the hardware more useful, and more hardware ships. Investors described the batch as more science-fiction than recent cohorts while saying valuations looked more grounded.
— Y Combinator · TechCrunch
🔗 Y Combinator: companies · 腾讯新闻: YC 最科幻的一届——AI 创业开始从模型走向电力、芯片和机器人
AI Daily Digest is published every morning by KD Agentic. Sources are linked inline; aggregator coverage is used for discovery only and is not cited as a primary reference.
Top comments (0)