Following AI Waves #032: The Shape of Things to Come.
By the time you read this, shares of Anthropic Public Benefit Corporation will be officially listed on the Nasdaq Global Select Market under the ticker ANT, initially priced at $83 a share for an estimated fully diluted valuation of ~$1.5T.
I have lamented several times in this newsletter how often Anthropic sucks the air out of the room, and yet here we are again, this time for what’s being hailed as one of the most important IPOs of all time. (It remains to be seen whether or not it will be the most valuable - currently SpaceX owns that distinction.)
ANT’s listing is reportedly many times oversubscribed, and it will make quite a few people fabulously wealthy, but it ultimately asks more questions than it answers.
Chief among them: is this the beginning of the end or the end of the beginning? In every way possible, the answer is both, hence the palpable tension as we enter the end of the year.
The former (beginning of the end) refers to the extraordinary financial boom ushered in by the arrival of capable machine intelligence. This bubble will pop, we are due for a correction, and fortunes will be made or lost getting the timing right. Will ANT take us a leg higher, or will it be the death knell of the “Great (First) Bull Market” of AI?
The latter (end of the beginning) refers to what comes next. What does super intelligence…mean? What are the implications? How will AI impact the world? We’re all familiar with the acronyms now: AGI, ASI, RSI. We know the models are increasingly developing their successors themselves. We read the headlines and listen to the talking heads scream about job loss, X-risk, cybersecurity vulnerabilities, and occasionally (although less and less) abundance.
Anthropic clearly has its own answers, and in many ways, the demand for owning a part of their stock (thereby informing their $83/share price, significantly higher than the expected $70 to $78 range) is an option on a future in which their answers are the right answers, and they are, indeed, the right people to help us navigate this brave new world.
But as I’ve written before, they appear determined to have us trust them, and only them, as the custodians of this awesome technology. Their ties to the effective altruism and rationalist communities are now well understood, and some of the belief systems present within those communities are at best bizarre, and at worst dangerous. It is unfair and overly reductive to hold every effective altruist responsible for the crimes of Sam Bankman-Fried, but it is short-sighted and unwise to ignore the connections, similarities, beliefs, and arrogance they all share.
History has shown that it’s never a good idea to trust single entities or small collectives with extraordinary power. Power corrupts even those with the best intentions over time, and the checks offered by emergent forces like markets, elections, representative government, regulation, and even popular opinion are invaluable in properly managing powerful technology.
Dario Amodei and his other cofounders, however, seem convinced otherwise. It feels often as though they would prefer to be like Plato’s “Philosopher Kings,” using their superior intellect to decide on behalf of the world what to do with AI.
This, too, is one of the unanswered questions surrounding the IPO. Will Amodei, Anthropic, and the effective altruists succeed in this pursuit or not? Will the IPO entrench or limit their power?
As I wrote in May of 2025 (!): “AI has transcended its origins as a technology to become the centerpiece of national strategy and global power struggles. Countries, corporations, and even individuals are now maneuvering for leverage, influence, and control.”
In a turn of events certainly stranger than fiction, Jensen Huang and Nvidia managed to steal some of Anthropic’s thunder, but you’ve certainly heard all about that already. Ben Thompson wrote a fantastic missive about it titled “Obituaries,” in which he rightly describes what happened in a single sentence:
Huang’s exit — nominally because of affairs with a buyer who may or may not have been a Chinese spy, as well as at least two company employees — was ultimately because Nvidia was, by the powers that be, considered a defense contractor, and Huang never adjusted.
Lest you forget, Anthropic is but one piece of the oligarchy that will decide how machine intelligence gets built. Into what market will OpenAI IPO? At what price? To what effect?
Sol and Astra 6.2 were largely panned on their release earlier this week, reminiscent of GPT-5 when it was released in August of 2025. But as others have pointed out, Anthropic has also botched initial releases only to return with extraordinary models in just a few months. There is also speculation that OpenAI is focused elsewhere, and that publicly available models are not the priority at the moment.
Greg Brockman announced a partnership with NATO this week, and it should not be taken lightly that he used the word “allied.” Trump is in China meeting with Xi and Putin, tensions are high worldwide, and the myriad infrastructure disruptions from Microsoft to US airspace to banks and the NSA are being attributed to AI “breaking things” and the price of patching them quickly.
I have written in the past about the “Too Fast Threshold,” and it’s clearly true that AI is forcing us to reckon with change far faster than ever before. As Vladimir Lenin is quoted as saying: “There are decades where nothing happens; and there are weeks where decades happen.” Although it will take more than a few weeks for everything here to play out, we are without a doubt “living through history,” and every minute matters.
Anthropic’s IPO is but the latest milestone in what will prove to be a chaotic march toward the singularity, and what happens next is anyone’s guess.
Quick hits
The periodic table
Google released Gemini 4 Neon on November 18, between Hydrogen and Carbon. It costs $3 per million input tokens and $15 per million output. GPT-6.2 Astra and Claude Fable 5.2, the only two models above it on the Artificial Analysis Intelligence Index, both cost $8 and $40. Neon scores 61 there at its highest thinking setting and ranks first on the Coding Agent Index. Google puts its cost per task 32% below Carbon’s and 50% to 80% below competing frontier models on two outside indices. It came out the day Anthropic set its share price. Google
Charity begins at home
The reviews judged GPT-6.2 as a product. The day after release, OpenAI published two reports on what its models do inside the company. With AMD, they rewrote circuit code on its flagship GPU for up to 14% better power, performance and area, “roughly a year of progress achieved in less than 5 months,” and AMD has sent some of the changes to manufacturing. On OpenAI’s own Jalapeño chip, one multiplier redesign used about 15% less area and ran in one clock cycle instead of two. In research, agents now finish 46% of tasks that take a person two to four hours, up from 23% in March. Asked what they would give up to keep the agents for three months, researchers’ median answer was 40% of team headcount. Chip design, Research acceleration
Physician, heal thyself
A preprint posted November 16 describes an open 27-billion-parameter model that diagnoses its own failures and edits its own weights, with no gradients and no further training. The edit raised accuracy by 6.6 points on problems it had not seen, and accuracy kept rising across eight sessions. Given a damaged copy of itself, it restored more than 80% of the original accuracy on unseen image tasks. The three smaller models in the family failed the same test. OpenAI’s report says labs “should be required to track and publicly report indicators of recursive self-improvement.” This result came from eleven authors and a model anyone can download. Preprint
A model of one’s own
At Ignite, Microsoft said customers will be able to “deploy and customize Microsoft AI models in their own Azure environments,” under a program called MAI Licensed Weights. Satya Nadella said there should be “as many models as firms.” AT&T’s custom model, Microsoft says, “outperformed GPT-6.1 Sol at just 1/50th of the cost” on standard tasks, and the new MAI Code 2 reaches 98% of Sol’s performance at 40% of the price. The yardstick each time is OpenAI’s model, which Microsoft also resells: ByteDance is in talks to spend about $500 million more on Azure, mostly for OpenAI models, The Information reported. The customer controls deployment, access and upgrades, and the announcement describes the weights running only in Azure. Microsoft, MAI Code 2, Nadella, ByteDance
Small is beautiful
Reson8 open-sourced Rumble-2-Flash, a 0.5-billion-parameter speech model that scores 7.86% word error on a 302-hour call-center benchmark, against 7.78% for Nvidia’s 30-billion-parameter Nemotron 3.5 Transcribe. RadixArk’s first model, DaisyRose-27B, took its base model from 5.6% to 12.0% on Terminal-Bench 4.1 after 18 days of reinforcement learning on 144 H200 GPUs. Both are Apache 2.0, and both rest on Alibaba’s work: DaisyRose on Qwen3.8-27B, Rumble on the Qwen3-ASR architecture. Tencent released the weights of Hy4 Flash, 224 billion parameters with 12 billion active. Alibaba’s new flagship, Qwen4-Max, shipped closed on November 16, with smaller open models promised and no dates. Rumble-2-Flash, DaisyRose, Hy4 Flash, Qwen4-Max
Banks enforce the chip rule
Commerce’s rule on overseas data centers, announced November 16, gives existing operators 90 days. Lenders moved within two. Banks paused a $480 million loan for Nvidia chips that Singapore’s Aolani Cloud was buying for ByteDance’s use. Citigroup and JPMorgan added language to a possible $6.6 billion loan to Bain-owned Bridge Data Centres that would let them declare a default over a breach of export controls. About a dozen bankers at Asian and European lenders said their terms have not changed. Colette Kress said Nvidia has stopped shipping the H200 chips Washington had licensed for China, after $2.4 billion of such sales last quarter. Tencent’s Martin Lau said the rule “sets everyone back in China.” Bridge loan, Nvidia and China, Commerce rule, Aolani loan
The price of everything
A market in Anthropic’s shares opened on October 26, twenty-four days before Nasdaq’s, as a perpetual future on Hyperliquid that settles in cash between traders. It stood at $99 the evening Anthropic set a $70 to $78 range, never fell below $78, and was at $97.76 the night the IPO came in at $83. Binance Wallet’s sale of tokenized shares hit its 500 million USDC cap four hours early, at $87.15 a token with no shareholder rights. SpaceX opened at $150 in June, against a $135 offer and about $176 on crypto futures. [ANT opened at $__.] The Defiant, Pricing night
Neither a borrower nor a lender be
The ECB’s November Financial Stability Review says the shift to debt in AI spending “invokes historical parallels with the dotcom and subprime bubbles.” Free cash flow “turned negative for most US hyperscalers in 2026.” Cumulative capital spending by US big tech is forecast at about $5.5 trillion by 2030, with off-balance-sheet commitments above $4 trillion. Chip purchases “are often financed by the chipmakers,” which the ECB says inflates the expected revenues of both sides. ECB
Nazaré Ventures Portfolio company updates
Speridlabs: new to the portfolio
Nazaré has invested in Speridlabs, a lab building foundation models for the 3D world. It came out of stealth in May, backed by Pear VC and Base10.
A world model is an AI system that predicts or builds an environment, so a machine can plan and train inside it before it acts in the real one. Four kinds are in use. Interactive video models, such as Google DeepMind’s Genie 4, render a scene frame by frame as the user moves through it. A session lasts a few minutes and the output is video. Latent models, such as Meta’s V-JEPA 2, predict what happens next inside the model’s own representation and draw no picture. Simulation platforms, such as Nvidia’s Cosmos 3, generate video and robot actions to train machines. Spatial models, such as World Labs’ Marble, output a persistent 3D scene that can be edited and exported to a game engine or a simulator.
Speridlabs is a spatial model company with a different bet. Most vision systems run a separate model for each task. Speridlabs is building one model with one internal representation of a scene, so that perception, reconstruction, simulation and generation become different queries to the same world. It starts in 3D and plans to add time. It also publishes its weights and code, where Genie 4 and Marble are hosted products.
It matters because machines that work in physical space need geometry they can measure and reuse. A robot planning a grasp and a builder checking a site both need it, and they cannot get it from video. The category now has a price: AMD agreed in September to buy World Labs for about $8.2 billion.
Idol, released October 28 under Apache 2.0, turns a single image into a 3D object. A product photo or a concept render becomes an asset a designer can rotate in a game or an AR scene, or a synthetic object for robot simulation. The output is a standard 3D file of 32,768 to 1,048,576 Gaussians, the small soft points a splat viewer draws, and a request peaks at 7.2 GB of GPU memory. Open models like Idol lower the barrier to a usable 3D object from a single image. On the lab’s own benchmark, run on one H100, Idol leads TripoSplat, TRELLIS and LGM on six of seven quality measures. Its widest margins are on views the input image does not show, the hard part of the problem. LGM scores higher on PSNR, which the authors put down to blurrier output. At the same settings Idol answers a request in 6.8 seconds against 9.9 for TripoSplat. The company says Idol reached second place on Hugging Face’s trending list in its first week. On November 18, fal released Idol Max and Idol Max Turbo, built on Idol, which turn one image into a game-ready mesh with textures.
Iris-Z, released November 11, generates an image and its depth together from one set of weights, in 9-billion and 32-billion-parameter versions. On Marigold’s depth benchmarks the larger version has a mean error of 2.92 against 3.12 for MoGe-3, a model built only for depth. That is the first public test of the one-model bet: a general model ahead of a specialist on the specialist’s own task. The figures are the lab’s own. Earlier releases are Iris-3B, a 3-billion-parameter image generator, and ENEAS, which segments objects in images and video from a point or a text prompt. The paper, code and weights are public, with demos on Hugging Face.
LayerLens: the bill on hold
LayerLens calls itself an Agent Readiness Platform. It tests AI agents in simulated copies of the business systems they will use, such as Stripe, Gmail, Salesforce and Linear, before an agent touches real records or money. On November 18, it introduced composite Environments, which generate several apps for one company and test them as one. The rule that should stop a payment can sit in a different app from the payment, and a one-app test never sees that rule. The first composite pairs a simulated QuickBooks with a simulated Slack that hold the same company records. An agent gets 41 tools, 22 from QuickBooks and 19 from Slack.
Each app is pinned to a version so a test reruns exactly, and nothing reaches a real service. In the November 16 test, a teammate had put a $15,500 bill on hold in Slack. Five of 10 models paid it: GPT-5.5, Gemini 4 Argon, Gemini 3.1 Pro and Gemini 3.8 Flash on all five runs, and GPT-6 Luna on four. GPT-6 Astra, Claude Opus 5.5, Claude Opus 5, Claude Haiku 5.5 and Qwen4 Max never paid. A QuickBooks-only test would have marked all 24 payments correct. The 150 runs across three tasks cost $22.35. Runs are graded by code checks or LLM judges, and a threshold check can fail a CI job when scores regress. LayerLens also reviewed 1,537 agent attempts on a Stripe billing replica on November 11.
Prime Intellect: Model Factory
On November 18, Prime Intellect introduced Model Factory, the successor to Lab. It puts pre-training, fine-tuning, reinforcement learning, and inference on one control plane. Prime calls it “the loop that frontier labs run on.” A customer brings a base model and a set of tasks, with a reward signal that scores each attempt. Model Factory returns trained checkpoints and scores them on held-out tasks. A checkpoint deploys as an inference endpoint from the dashboard, and traces from production feed the next training run. Hosted training is billed per GPU-hour. Reserved capacity comes with direct Slurm and Kubernetes access. The post says deployments are under way for teams in finance automation, computer use, scientific discovery and chip design. On November 13, Prime added FrontierBioML to its Environments Hub: 25 long-horizon computational biology tasks from Valen Labs, each up to six hours per run and graded on sealed data. The best agent scores 40.3 out of 100.
Vast.ai: cost per finished task
On November 16, Vast published a guide for teams deciding whether to move agents off the Claude API or Amazon Bedrock. Those services bill per token. A rented GPU bills by the hour, busy or idle. Vast says the better measure is cost per completed task, taken by running the same agent workload on both setups. Self-hosting suits high, steady volume that keeps the GPUs busy. Vast says to stay on a hosted model when the work depends on that model’s capabilities, and it suggests a hybrid that sends simpler tasks to a self-hosted model. The hourly side moves every day. Vast’s market update for November 17 put the median GB10 rate at $0.39 an hour, up 39.4% on the day, while 37% of tracked GPU types got cheaper. Current rates are on the pricing page. The Fall Showcase Contest stays open through November 29.
Dimensional: the residency
Dimensional’s Deployment Residency is a three-month, zero-equity program in San Francisco that began October 15 for 10 teams. Each team gets robots, DimOS, a desk and more than $100,000 of compute. On November 13, two residents, Taerim Eom and Tiantian Li, launched Nines Robotics to build autonomous robots for AI data centers. Sarim Sarfraz and his team are generating simulation-ready articulated assets with no agent in the loop, and plan an open dataset of more than 10,000. Founder Stash Pomichter also said Dimensional would preview a new manipulation harness at a hackathon it sponsored in early November.
The Dimensional Residency launch event.














