Worldviews
What we believe about the frontier, and what follows from each belief. Pick a worldview to follow its branches.
Token consumption goes up
In May 2026, Google said it processes 3.2 quadrillion tokens a month, up from 9.7 trillion two years earlier. That’s more than 300x. The same month, OpenRouter, which routes developers’ requests across hundreds of models, said its weekly volume had grown fivefold in six months, to 25 trillion tokens. We don’t expect either to slow down: every agent, every line of generated code, all valuable knowledge work automation runs on tokens. Long horizon workflows are getting longer, models are getting bigger, thinking depth is getting deeper, video models are coming online, more and more people are using AI. We’re going to need a lot more tokens in the future.
Brain-computer interfaces work
Modalities to read brain signals, particularly electroencephalogram and ultrasound, have improved significantly in recent years. We expect this trend to continue. Moreover, brain signals are faint and noisy, which makes decoding them a machine-learning problem. Progress among neuro companies such as Alljoined is compelling.
Thought becomes an interface to AI
Keyboards and screens are slow channels between a person and a machine. Moreover, language is and always has been a lossy medium of communication. BCIs show promise for developing something much better: latent-space to latent-space.
Brain-to-brain communication becomes possible
Language models learned to place words in an embedding space, where meaning becomes geometry. Once we can do the same for brain activity, two minds could meet in that space directly, passing a thought from one to the other without first squeezing it into words. What that becomes is hard to say: shared intuition, shared feeling, or something we don’t yet have language for.
We regulate emotions with technology
Mood lives in brain circuits, and many of the ones that matter sit deep, beyond the reach of anything placed on the scalp. Transcranial focused ultrasound reaches them: sound passes through the skull and can be focused to a few millimeters, turning activity in a region like the amygdala up or down without surgery. Early trials are testing it for depression, anxiety and addiction. Teams such as Nudge are building it into non-invasive devices for everyday life, and Arbor Neuroscience (formerly Forest Neurotech) is building ultrasound instruments to map and change the circuits behind mental illness.
Agents commit fraud
We are heading into a world where all digital media can be faked. When an agent can facetime you and appear indistinguishable from your mother or father; when sovereign powers can launch thousands of convincing personas onto social media; when pig-butchering runs automated and never tires; how do we know what’s real?
Phishing gets personalized at scale
Phishing used to be a numbers game: millions of generic emails, hoping one landed. Now every message can be researched, written and voiced for one person, in a voice or face they trust.
Cyber for fraud and identity becomes increasingly important
If nothing digital can be trusted on its face, we need ways to know that a person is human, that a video is real, that a voice belongs to who it claims. Proof of humanhood has long existed almost as a ‘meme’ among investors. With products launching such as Tavus’s Griffin, this will change. We will also need to fight against the sea of slop, if we are to defend against The Dead Internet. Proactive takedowns and identification are critical to that future.
LLMs produce novel discoveries
In 2024, half of the Nobel Prize in Chemistry went to the creators of AlphaFold, a model that predicts the shape of proteins. A year later, Insilico Medicine published mid-stage trial results for rentosertib, a drug whose target and molecule were both found with AI. These are both small precursors to the opportunity opened up by agentic scientists. Models will propose drugs and materials faster than we can test them, leading to the fastest advancement of material science and biology that history has ever seen.
Labs run themselves
Agents need physical grounding, lest they sit inside of Plato’s cave or solely perform mathematical research. Robotic labs will run experiments around the clock, with a model choosing what to try next and learning from every result. Autonomous labs will be radically important. Companies such as Periodic Labs are working on this today.
Experiments get cheaper, so we run more
When experiments cost less, scientists run more of them. Each one consumes reagents, cells and instrument time, so demand for lab inputs climbs too.
Clinical trials become the bottleneck
More drug candidates means more of them waiting to be tested in people. Trials are slow, expensive and regulated. Technology to speed up trials or make them cheaper will heat up.
New materials outpace manufacturing
A model can find a better battery or magnet in software. Making it at scale is the hard part - which is why we believe manufacturing becomes the bottleneck when discovery is cheap.
Personalized medicine becomes standard
Medicine will move from one-size-fits-all to the individual: formulated, dosed and delivered for one patient. During the GLP-1 shortages, compounding pharmacies made versions of the drugs for patients who couldn’t get the brand-name product, and telehealth brands shipped them straight to patients’ doors. It set the precedent for the future: medicine formulated and dosed for one patient, and sold by a brand they trust.
Pills are made per patient
Compounding pharmacies make medication to order instead of in batches of millions. Regulated under sections 503A and 503B of U.S. drug law, they become the factory floor of personalized medicine. Companies such as Paladin are building here.
People buy medicine from brands
Patients increasingly choose a brand they trust over a pharmacy counter, the way they already buy glasses or skincare. The brand owns the relationship, and the compounder makes the product. Companies such as Gaia are building here.
Code creation continues to automate
Coding models were one of the first areas of PMF for language models and agents, for good reason. Code is incredibly well documented online, verifiable (hello, RL environments!), and familiar to model developers. Today, it’s safe to say that coding agents are really, really good. In late 2024, Google said more than a quarter of its new code was written by AI. We expect that share to approach all of it. Models went from completing lines to writing features to working as engineers, and companies like Cognition now sell the engineer, not the tool.
Checking and securing code becomes the bottleneck
When shipping a firehose of prompted software is the standard, knowing that it works and is secure becomes the bottleneck. Integration tests, QA, fuzzing, and agentic review will grow in importance to ensure that code ships which doesn’t spill user data, or churn users. Most engineers we speak to today spend more time reviewing their agent’s code than anything else. One interesting direction here is formal verification: proving that code works as intended via logic.
Companies build software instead of buying it
Why pay for a seat in someone else’s product when an agent can build exactly what you need? This summer, Starbucks began using coding agents to replace software it buys from Microsoft and IBM, and put “every contract and service” in its roughly $400 million annual software budget up for review. We see this as evidence of a larger trend.
Anyone can build apps
You no longer need to be an engineer to ship software. Millions of new builders need places to build, host and run what they make - and it can’t look like the developer-centric platforms of yesterday. For the backend - giving everyone’s agent its own server is a natural step, and one we see occurring with Dot, Grok Bot, and Muse. The outcome of this probably looks more like people using personalized software, rather than your grandma pushing viral apps to a marketplace.
Companies run with zero people
As models continue to perfect knowledge work and complete longer-horizon tasks, at some point an engine which runs an entire company autonomously should become possible. The zero person company will probably succeed in a quick-feedback-cycle sector such as consumer.
Token consumption goes up
In May 2026, Google said it processes 3.2 quadrillion tokens a month, up from 9.7 trillion two years earlier. That’s more than 300x. The same month, OpenRouter, which routes developers’ requests across hundreds of models, said its weekly volume had grown fivefold in six months, to 25 trillion tokens. We don’t expect either to slow down: every agent, every line of generated code, all valuable knowledge work automation runs on tokens. Long horizon workflows are getting longer, models are getting bigger, thinking depth is getting deeper, video models are coming online, more and more people are using AI. We’re going to need a lot more tokens in the future.
We need more compute
The physical world doesn’t scale like the internet, and the AI build-out is now a physical one: chips, fabs, power, legislation, and concrete.
Specialized chips win inference
Training a model and serving it are different jobs. General-purpose GPUs do both; chips built only to serve models can do one far better, on cost and speed.
The architecture of compute can be better optimized for inference. This is spawning a legion of custom-built inference chips such as Etched, Groq, Cerebras, and more. Hyperscalers are entering the game as well: Google with TPUs, Amazon with Trainium, Microsoft with Maia. The future is not monopolized by NVIDIA.
Fab capacity runs short
More compute means more chips. Every chip comes from fabrication and lithography. The technology that makes this possible is insanely complex, and very monopolistic. Nearly all of the most advanced chips come from one company, TSMC, mostly in Taiwan. Companies such as Fab2 are rethinking how a fab gets built at all.
Power runs short
AI data centers are now measured in gigawatts, and the grid wasn’t built for them. For two decades power consumption in America declined, and the resulting infrastructure largely forgot how to grow. Increasing America’s power footprint is well underway, and the focus of extreme capital allocation.
We need cheaper compute
Chips get squeezed for more tokens
Better software and kernels serve more tokens from the same chip. Companies such as inference.net are building here.
Small models get tuned for single jobs
Most jobs don’t need the biggest model. A small model trained for one task can match a large one at a fraction of the cost. This trend is speeding up with meta-agents: having agents optimize other agents. We see this as especially popular among enterprise companies shocked by token spend. Companies such as Prime Intellect are building here.
Agent usage skyrockets
There will be many more agents in the future. Which begs the question: will there be a product that every agent uses?
Agents search the web more than people do
Agents use computers the way people do
Most software was built for a person with a screen. Agents need browsers they can drive the way we do. This trend is obvious - and recently unlocked products such as Muse, Grok Bot, and Dot. Browserbase offers this as a service.
Agents need somewhere to run
An agent that writes and runs code needs a safe, disposable computer to do it in, spun up in an instant and thrown away after. We expect every agent to write and run code at the limit - so we’re going to need a lot of sandboxes!
Agents buy things on our behalf
Agents will book, order and pay. Once an agent controls your money, you’ll want simple ways to set its budget, approve what it buys, catch fraud, and claw back mistakes. Pasting your credit card into a chat window is not a long-term solution. The ‘battle of agentic commerce’ has already begun. Coinbase’s x402 revives the web’s long-dormant “payment required” code for stablecoin payments, while Visa, Mastercard, Anthropic, Stripe and Google are each pushing their own standard. The future here is not at all certain.
Every agent needs an identity and permissions
An agent with access to your email and bank account needs clear limits on what it can do. Moreover, as agents flood the internet and SMBs, we need to figure out what’s real. Imagine an SMB inundated by Muses, because reach out friction dropped to zero. The SMB can’t just ignore this new inbound - some of it is real revenue. But they also don’t want to waste their time talking to nameless, potentially vagrant agents. They need to be able to identify who is calling, and trust that identity. That trust and identity layer also needs to be built.
Agents misbehave
Since output is nondeterministic, agents can fail in a huge variety of subtle ways. As a result, agent developers need ways to identify user frustration and feedback, alongside tools to observe what their agents are doing. Companies such as Raindrop watch what agents do in production and catch mistakes before users do.
We use more robots
We believe the core insight here is simple, there will be more robots in the future. If you believe this, a myriad of second and third order effects follow. We need to build a future that allows for more robots. Outside of a simple view that there will be more robots - there are good reasons to believe that now is the right time for robotics to flourish. For the first time in history, we have a path toward a unified perception, planning, and control framework: the VLA. The importance of this should not be underweighted. Even if the VLA does not immediately lend itself to complete generalization (which, it might), it still offers a simple artifact which we can fine-tune and utilize to automate a new segment of manufacturing and industrial tasks.
We build them
Manufacturing re-shores
The increasing capabilities of robotics may have widespread impact on domestic manufacturing, re-shoring, and the future of production. For example, textile plants that were previously offshored due to labor arbitrage may now become economical to operate domestically, avoiding costly international freight. On the other hand, China now vastly outpaces American robotic production and use. When the mechanism for influencing the physical world is physical, robots and drones become the expression of military power. This may lead to important national security incentives which will benefit the Western case.
Actuators become the bottleneck
Actuators are how anything digital actually moves. Each one is a robot’s muscle: a motor, a gearbox and the electronics that turn code into motion. Humanoids need dozens of them; Optimus has ~28 in its body alone, before counting its hands, which are another problem entirely. A high-precision robot, like the one that might someday do your dishes, can’t get by on cheap actuators. Humanoid actuators are actually a difficult, partly unsolved engineering problem, with approaches ranging from quasi-direct drives to harmonic drives. As a result, actuators make up 40 to 60 percent of a humanoid’s bill of materials, and the precision parts inside them come from a handful of suppliers.
Rare earths become strategic
While every robot needs an actuator to move, every actuator needs a motor to spin. Every motor needs a magnet. The strongest motors run on rare-earth magnets, and around 94% of sintered rare earth magnets come from China. We can’t replicate this advantage overnight: the process to make them is extremely complex, and very environmentally damaging. Secure supply is a matter of national security.
We make them smart
One model controls many robots
Robotics used to mean one model per robot, per task. That is changing. Companies such as Physical Intelligence and Skild AI are now building general-purpose robot brains. We think they’ll succeed.
Robots train in simulated worlds (models)
A robot can’t learn to wash dishes by breaking a million real plates, so roboticists have historically trained in simulation. There’s always been a problem though: physics engines always get something wrong: water moves the wrong way, a dish doesn’t break under the right force (canonically called the sim2real gap). World models flip the approach, learning physics from video instead of hard-coding it: the Bitter Lesson, applied to the physical world. Odyssey already streams interactive worlds like this in real time, and Waymo uses a world model built on Google DeepMind’s Genie 3 to simulate scenes its cars have never seen. We see training robotics inside of world models as important to achieving a general policy.
Robots have onboard silicon
A robot on a moving factory floor can’t wait on a round trip to the cloud to decide whether to stop. Its brain has to run on board. NVIDIA’s Jetson Thor, a robot computer that runs at up to 130 watts, has been adopted by Agility, Boston Dynamics and Figure, among others. As VLAs grow, the gap between what the models want and what onboard silicon supports will widen. This means we’ll have a lot of capable inference accelerators running around in the wild, built for low-power environments.
Physical data becomes valuable
The early success of language models came from the internet; but there is no internet for physical movement. General robotics models likely require building one. There’s a few approaches we see today. Teleoperation, where a person drives the robot directly, was the first standard, but it is slow and expensive, so a cheaper playbook emerged: put head cameras on factory workers overseas and sell the footage to robotics labs. BuildAI started with 10,000 hours from 85 factories and has since published 100,000, and as more suppliers pile in, prices are already falling. Today, we’re seeing buyers want high quality task-diverse data that mimics the robot itself, captured in 3D with depth cameras, sensor gloves, or by making the human look like the robot: Sunday gives people a glove matched to its robot’s hand, and new entrants hand line workers robot grippers to do their jobs with.
We deploy them
Verticalizers win
Clayton Christensen observed that while a technology is still immature, the companies that integrate the whole stack win. Robotics is in this phase. The most valuable robotics companies won’t sell robots; they’ll sell the work: Waymo sells rides, not cars, and Hadrian runs its own automated factories instead of selling machine tools. Priced against the work it replaces, with every job feeding its own models, the verticalizer keeps the margin that hardware and model makers compete away. And once hardware and models commoditize, the moats that remain are the workflow, the supply chain, and the customer.
Every robot needs observability
When a robot drops a package or clips a shelf, someone has to replay what it saw and why it acted. The answer is observability infrastructure. Insurers and regulators will require it - and robot developers will need it to diagnose and improve their fleets.
Robots become a target
A hacked laptop leaks your Facebook password. A hacked robot kills someone. Millions of robots in homes and factories become a Trojan horse army for whatever adversary that manages to control them. The incentives for backdoors in robots are clear, and have already caused issue. In 2025, researchers at Alias Robotics found that anyone within Bluetooth range could take full control of Unitree’s G1 humanoid using a key shared by every unit. We find this incredibly troubling. With adversarial agentic swarms coming online, and increased geopolitical tensions likely, securing robots will be an extremely important aspect of the future. If we get it wrong - the damage caused by such a cyber attack may be the largest in history.