It feels small to some people in 2026, depending on the business type and the field. Most businesses are small businesses, though. It actually approaches 100%. This is under the SBA's definition as of not long ago. There are limits on earnings, number of employees, or both for what can even be considered a "small business" that vary by industry. They are as low as $9 million revenue or 100 employees, although some industries have much higher ceilings. https://legacy.sba.gov/sites/default/files/2023-06/Table%20o...
The average annual revenue of a small business in the US in 2026 is around $1.2 million. The average headcount is about 11 people. About 46% of employees in the US are employed by small businesses. In 2024, about 64% of new jobs were created by small businesses. About 61% of new jobs from 1991 to 2025 were created in small businesses.
Well, commodities are commodities, and it was said it was a niche market. The company itself seems pretty big to me, relatively. $400m in a market of about $2b means they fill a pretty substantial part of the available market.
I don't recall that I ever did this with a FoxPro app. I used to make part of my money porting or cloning bespoke small business apps to OpenOffice Basic or OpenOffice with its Python integration. Visual Basic, QuickBasic, MS Basic PDS, TurboBasic, MS Access, and Turbo Pascal apps I'd replace with the integrated languages of the office suite.
Some of those originals used DBase III or DBase IV libraries. Some used bdb libraries. An awful lot used fixed-width fields designed into the application itself. Many of them I wouldn't have the source for, so I'd have to reverse engineer the data file format a little and then write a proper schema for a real database.
Sometimes the hardest part was getting the forms to look acceptable to employees at my client who were used to a curses style interface.
If we're all using distilled open-weight models in ASICs in our own systems the energy cost will come way down. The question is when that becomes a reasonable solution for a broad set of use cases.
A MSTR only needs uranium-233, uranium-235, or plutonium to start. It then produces uranium-233 as part of its fuel cycle as thorium is input.
Given current found reserves and the current rate of use, the world has about 40 to 50 years of natural gas. Thorium used in molten salt thorium reactors would provide electricity for 60 billion years or so if we could actually extract all of it. That's 10 billion years or more if it provided all human energy consumption. Of course there's a limit to extraction, but it is over three times as common as uranium.
Also, besides thorium one can mix in partial amounts of other fuels, including uranium and plutonium. There is no runaway meltdown risk, as the fission is actively managed rather than actively suppressed. Fuel is spent more completely. The waste products are smaller, less radioactive, and have far shorter half-lives.
Then of course we're always getting slightly closer to productive fusion reactors.
These technologies along with solar PV, solar thermal, hydro, wind, geothermal, wave power, and batteries likely all have a place in the future.
There's a decent chance that at some point in the future residential customers will pay for the connection and only commercial or industrial customers will actually be metered. That's not because companies want to give up additional revenue. It's because at some point the cost of meters, tracking usage, and competitive advertising about who has the cheapest plans costs more than the power the typical customer uses above the base charge.
It contains a different mix of elements and isotopes than spent solid fuel rods from light water reactors.
The reactor creates new fissile material while it runs, then fissions that. The new fissile material is recycled into the reactor. Even actinides can largely be recycled into the fuel stream. The fuel and coolant being a liquid mixture allows a lot of chemical processing and returning parts of the initial waste back into the reactor more fully.
The waste that's actually handled for storage tends to have a lower proportion of fission byproducts that carry their own radioactivity, but some traces of highly active sources. I've read cesium, strontium, iodine, xenon, krypton, barium, and various noble metals are the bulk of the waste. Some of that cesium will be cesium-137. Some of the strontium will be strontium-90.
Solid rods from a light water reactor are not even nearly completely spent before they become too degraded for producing electricity. They contain cesium-137 and strontium-90 too but along with uranium-235, plutonium-239, americium-241, neptunium237, and curium at the time they're ready for storage. Small parts of this can be reprocessed rather than stored, but often for nuclear weapons. Having solid rods also makes it more difficult to separate elements.
MSTR waste is dangerous, but for about three hundred years. The radiation involved tends to be largely gamma, but from a very small portion of the waste rather than the bulk of the mass. Light water reactor waste can be dangerous for tens of thousands of years, producing alpha, beta, and gamma from most of the mass of the waste. The alpha and beta radiation will last tens of millennia.
I guess if you abused terminology to scale by risk then, something decaying faster but by a different mechandism could be “less radioactive”, e.g., for external exposure, an alpha emitter with a shorter half-life might be “less (dangerously) radioactive” than a beta emitter (of course, the reverse would be true for internal exposure.)
Can’t really think of any other way to rationalize that combination.
What about inference on a distilled version of someone else's model? What about inference on an open weights model? I don't think all LLM business models are going to revolve around developing state of the art frontier models.
Right. This is what I was implying in my comment, but it's good to be explicit about it. I think there will be successful companies that just sell inference against the best models they can get without needing to invest anything (or very little) in training anything new. This could even be a spinoff of one of the big frontier labs. I would personally rather invest in an IPO for a company that only owns the gpt-6-astra implementation and infrastructure than in openai itself. There is certainly less upside, but IMO also way less downside risk. (I'm not saying there is any chance this kind of a spin-off would happen, it's just a thought experiment.) And I think this same calculation applies to open weights inference providers.
Edit to add: Or or might just be AWS / GCP / Azure that benefits from this business model. They're already pretty good at selling commodity infrastructure.
reply