If we succeed, we will enable growth across the entire economy to accelerate, by letting enterprises grow AI on their own terms and letting more new labs train frontier models.
The short version: AI is just getting started, but its benefits are dangerously concentrated. Today, the progress is disproportionately captured by a handful of industries, most visibly software engineering and by a handful of labs with the compute, talent and capital to train frontier models. Meanwhile two groups are largely locked out: the companies that hold the world's most valuable domain knowledge and proprietary data, such as banks, hospitals, manufacturers, governments and law firms, and the new labs that want to train frontier models but don't own a hyperscaler's infrastructure. The enterprises can rent someone else's general-purpose model, but they can't easily infuse it with what makes them unique. The labs can rent raw compute, but they inherit vendor lock-in, low utilization and a software monopoly they didn't choose. We think closing both gaps is the single biggest untapped lever for economic growth this decade and it's what we're building SF Tensor to unlock. The same stack that lets an enterprise post-train models on its own data lets a lab pre-train a novel architecture from scratch, both running across multiple hardware vendors at state-of-the-art utilization, reducing training costs by an average of 70%.
Our Seed round was led by Susa Ventures, with participation from Y Combinator, BoxGroup, Soma Capital, Liquid2, Transpose Platform and a group of angels including Max Mullen, Paul Graham as well as founders and executives at NeuraLink, AMD and Notion. We're using this funding to put AI-building capability into the hands of the people who actually run the economy and the people building its next models.
Our belief in the future of AI
We firmly believe that we are just seeing the beginning of AI. While progress on general-purpose text models is starting to reach a point where there are debates about whether it's slowing down, the far larger story, the one that will actually move GDP, is barely underway: the diffusion of AI into every enterprise, trained on the data and expertise that those enterprises already own.
Right now, the value of AI is disproportionately concentrated. It accrues to the frontier labs and to the industries, chiefly software engineering, whose work happens to look most like the public internet text these models were trained on. That's not a coincidence. A model is only as good as its exposure to a domain and the open web is saturated with code, forums and documentation about building software. It is comparatively starved of a hospital's clinical protocols, a bank's decades of underwriting decisions or a manufacturer's process telemetry. So the enterprises sitting on the most valuable data in the world get the least benefit from the current paradigm, precisely because their edge is the thing the public models never saw.
Closing that gap is the point of this company.
Enterprise models: post-training on what you already own
We see two major categories of enterprise AI and both depend on the same missing capability: the ability to cheaply and repeatedly train models on your own data.
The first is models trained on specialized or proprietary data. Specialized data is generally public or quasi-public, such as GPU kernels, that frontier models could be trained on but simply aren't, because it makes no sense for the vast majority of use cases and the training compute is better spent elsewhere. But for the organization that lives in that domain, that data is the entire game. The bigger category is proprietary data: a government agency's prior reports, a hedge fund's internal analysis, an insurer's claims history. Handing this to a frontier lab and risking it leaking through a public model, is a non-starter for most serious enterprises.
For these cases we use what we call "data injection training": we take a model trained roughly 80% of the way, inject a company's custom data, then complete the remaining 20% including post-training, RLHF and so on. This lets the model deeply integrate proprietary knowledge rather than bolting it on at inference time and the process can be repeated continuously as policies, regulations and business conditions evolve. For a smaller set of customers with the scale to justify it, we support full pre-training from scratch, which is especially valuable in cases where new modalities are necessarily, such as drug discovery. But for the vast majority, the winning move is post-training and post-training is exactly what the current ecosystem makes needlessly hard and expensive.
The second is lightweight and specialized models, such as those deployed on-edge. These can't be as generalized as large-scale models, simply due to compression limits: the memorization and fact-storage aspect of a model is in essence a form of data compression, bounded by parameter count. (This is one of the primary reasons MoE models are smarter and better at memorization than dense models with the same number of active parameters.) But for a specific task, where general knowledge like the exact dates of global events is irrelevant and what you need is general reasoning plus domain-specific data such as sensor readings, audio or visual data, a custom model is both smaller and better. Enterprises will increasingly want fleets of these. It is our firm belief that there should be at least as many models in the world as there are companies.
The through-line is simple: the highest-value AI of the next decade won't be one more general model in the cloud. It will be tens of thousands of enterprise-specific models, continuously trained and retrained on data their owners can't and won't hand to anyone else. That only happens if training on your own data becomes cheap, fast and self-serve. Today it is none of those things.
Why the economic impact is so large
This is not a niche. The industries that have barely been touched by AI: healthcare, financial services, manufacturing, logistics, energy, government, are the ones that make up the overwhelming majority of GDP. Software is enormous culturally and in market cap, but it is a small slice of total economic output. If AI's productivity gains stay concentrated in software, the macro effect is real but bounded. If those same gains diffuse into the sectors that actually constitute the economy, the effect compounds across the whole thing.
That diffusion is gated almost entirely by one question: can a domain expert who is not an ML infrastructure specialist get a model trained on their own data, into production, affordably and keep it current? Every month that answer stays "no," the benefits of AI stay locked in the same few places. Our entire thesis is that making the answer "yes" is both the biggest technical opportunity and the biggest economic one.
From Ownership to Outcomes
Every major resource follows the same arc. First you own it: factories built their own power plants, companies racked their own servers, and until recently, if you wanted GPUs, you bought them and ran them in your own building. Then you rent it, but rigidly: long-term leases, fixed capacity, multi-year commitments to specific hardware. And eventually it becomes something you simply draw on. Nobody wiring a building today asks which power plant their electricity comes from. Nobody deploying an app on modern infrastructure specifies the chip architecture, the memory or the data center. You say "run this" and the rest disappears.
GPU compute is the last major resource still stuck at the beginning of that arc, and it's stuck harder than the rest. With CPUs, portability was never really the problem: run your code on Arm, Intel or AMD and it just runs. GPUs aren't like that. Move a workload to a different chip, a different cluster, a different topology and you don't simply recompile, you re-engineer. Months of hand-tuning kernels to claw back the performance you had before. So companies pick one setup and stay locked to it, no matter what the market does, no matter what's cheaper, faster or newer, because the cost of moving is measured in months of research velocity, not just dollars.
That lock-in is a real prison, and breaking it is the hard part. Our automatic kernel optimization and search compiler does the re-engineering that teams do by hand today, so a workload can move to whatever hardware, cluster or topology actually serves it best and run there at full performance, with no human in the loop. Once moving becomes free, hardware and TFLOPs go back to being what they should have been all along: a tool. You shouldn't buy GPUs, you should buy outcomes. "Run this training job, give me the results in 48 hours." What runs it, where it runs and how it scales is our problem, not yours.
But buying outcomes instead of hardware only works if the supply underneath it is actually open. And today it isn't.
The Compute Barrier Problem
Enabling that requires solving the compute barrier underneath it, because you cannot scale out model-building on top of a supply chain that only serves a handful of the largest buyers.
The barrier breaks into three parts: physical chips in limited supply, energy to power them in limited supply and a lack of competition in GPUs. Parts (1) and (2) are, to a large extent, an indirect result of part (3).
There are enough lithography machines in the world to produce far more chips than we have today. The problem is that everyone is chasing the latest generation of NVIDIA hardware, which requires a very specific node, a very specific supply chain and a single manufacturer, constrained by factors including CoWoS packaging capacity and TSMC's aversion to further customer concentration.
That does not need to be the case. Today there are multiple manufacturers producing chips competitive with NVIDIA on the raw hardware capabilities relevant to AI training. AMD's Instinct MI300X and MI355X offer competitive training throughput and significantly more memory. Google's TPU since v5 have powered some of the most impressive AI systems in the world. AWS has Trainium, with hundreds of thousands of Trainium chips powering Anthropic's models. Intel has its Gaudi accelerators. Cerebras has built wafer-scale chips with 4 trillion transistors and 900,000 AI cores. And startups like Groq, SambaNova and Tenstorrent are all bringing competitive hardware to market. Interestingly, all of the established players here are American 🇺🇸. The problem exists at the software layer and is further amplified by hardware availability.
When a company trains a model, especially one with limited resources, it must decide what hardware to target, because it is not currently feasible to target multiple chips simultaneously off the bat. That decision comes down to three factors: how many people can actually program the hardware well and how good the surrounding community and knowledge are, how mature and stable the software stack is and how available the chips are on the market. On all three, NVIDIA wins in 9 out of 10 cases, specifically, right now, the H100. Not because the B200 isn't a better chip, but because the Blackwell generation has been plagued with issues: design flaws that required a re-spin at TSMC, pushing large-scale production back by months; GB200 rack shipments delayed by overheating, liquid cooling leaks, inter-chip connectivity problems and software bugs that took months of supply-chain collaboration to resolve. On top of that, there simply isn't enough experience running these newer chips at scale, so many companies keep betting on the H100 as a dependable variable.
Look at the history of the GPU and CPU and you see real generational leaps. Recently, that has stopped. The H100 has an 814 mm² die, already pushing against the reticle limit of what a single chip can be. The B200 gets around this by using two dies connected via a 10 TB/s link, effectively doubling the silicon area. Across the chips released between 2022 and 2025, from the H100, MI300X, TPU v5, H200 to the B200, MI355x and TPU v7, all have roughly the same TFLOPS/mm². Over the 2 years between the H100 and B200, that metric improved only about 15%. GPUs haven't really gotten better; they've gotten bigger.
That has consequences for energy and cooling, but that's a question for another essay.
Given that the raw capability is there for many manufacturers and there are numerous startups ready to compete on specific workloads, it is clear the NVIDIA monopoly is one created primarily by software support, not superior hardware. And that monopoly is precisely what keeps enterprise model-building out of reach: it concentrates supply, keeps prices high and forces every buyer into the same queue.
To unclog this bottleneck and let enterprises loose on their own data and researchers loose on their ideas, we need to (1) target multiple vendors easily, (2) run hardware from various vendors in a stable, heterogeneous manner at frontier scale and (3) ensure availability of that hardware to the people who need it.
SF Tensor has two products to do this: the core SF Tensor stack, which takes code that is easy to write (PyTorch, JAX or our own Emma) and targets multiple vendors at state-of-the-art utilization; and the Elastic Cloud, which targets different vendors across different clouds at frontier training scale, treating everything in a stable, heterogeneous manner.
Together they let you drastically increase compute availability and cut the price of training runs by an average of 70%. Just as importantly, they create competition. If researchers and enterprises are no longer forced onto a single cloud or vendor, then vendors and clouds have to specialize and innovate, because if they are the cheapest or best option for a given workload we will automatically target them, with no one rewriting their software. A chip or cloud that becomes uncompetitive can no longer coast on lock-in; it has to win on the merits.
That is why we target the abstraction layer we do. It lets us swap the compute backend, both the hardware and the provider, with a single config parameter, always targeting the best option for a given workload at any point in time.
The Velocity Problem
Talking to startups, labs and increasingly enterprise teams that have gone from research idea to first medium-size model to production model, the vast majority of the effort goes to non-research work: managing infrastructure, negotiating hardware deals, writing and optimizing GPU kernels and finding the uniquely rare and expensive people who can scale models. The alternative is to train at very low GPU utilization and eat the cost. If a GPU runs at 7% utilization, it is 7x more expensive to train than if the code is optimized to 49%. With training runs regularly costing hundreds of thousands to tens of millions of dollars, that is unacceptable and for an enterprise without an ML infra team, it is often disqualifying.
Our stack takes easy-to-write code, which domain experts can produce at high velocity and compiles it to high utilization. A researcher can go from a local experiment on 1 GPU to roughly 64 GPUs by changing a config parameter and start training instantly. When it's time for frontier scale, that goes up to 10,000. Our compiler spends a few hours optimizing the problem and a frontier training run starts in hours instead of weeks or months.
This is the difference between an enterprise that can only consume someone else's AI and one that can build and continuously improve its own.
The Compute Access Problem
The final problem is getting compute to the people who need it. At the microscale, that means letting a specific team get the compute they need without negotiating dozens of deals over months and scaling it up and down on demand, since training demand can look like 16 GPUs, 20, 100, 16, then 10,000 for a few weeks, then back to 16, a curve almost impossible to accommodate today. Our Elastic Cloud lets a user simply choose the GPUs and amount they need; the system finds them, allocates them, manages the run and returns results or lets the user connect in and experiment and debug as easily as they would locally. At the macroscale, it means creating enough availability, through more competition, that the people working on valuable problems can actually get in the door.
This is why we are launching a new research grant so that teams working on novel problems can receive initial grants for experiments, with follow-up funding based on results. We're also launching a new series of rigorously benchmarked AI hackathons to get compute into the hands of people building on new and novel problems. More on both soon.
We also believe we can't predict the future. We don't claim to know exactly how enterprise AI will be built or how people will need to access compute. So we make compute accessible through as many channels as possible: from our web app (Elastic Cloud) to a simple API to agent-native protocols that let autonomous systems acquire compute on demand. If you need access in a way we don't support yet, reach out. We'll probably find a way to add it in a matter of days.
Thanks for reading this entire essay on our belief in the future of AI, how we hope to help spread its benefits across the whole economy and thanks again on behalf of Tom, Luk and me to our amazing investors and team for backing our vision.
