Why not 20x $10M grants for smaller companies? They're gonna throw this money with no oversight anyways so why not bolster the actual startup scene instead of a bunch of incumbents who all have more than enough cash? $10M could keep a startup running for 1+ years at its most crucial time. That's 10 solutions instead of 1 -- statistically one of them will be a massive breakthrough?
Who are these smaller companies, and what do they have to offer that these 4 don't? Chances are that the smaller companies themselves are licensing the LLM from Google/Anthropic/OpenAI, so why pay middlemen for no reason?
You’re telling me that you can’t find 10 worthwhile AI startups to give money to? I bet there are 1000 on crunchbase right now. With $10M some of them could buy hardware to build their own systems.
Maybe 20 10M contracts is a bit too small, but this point is completely wrong too. Part of the purpose of public expenditure is to promote healthy businesses, not just procure things are efficiently as possible
This is people's money, and people benefit from competition in the market
IMO if the government needs widgets it should ask for bids and contract based on price. Some broad restrictions are OK (e.g., prefer domestic manufacturers), but government has no business skewing the contract by promoting <insert favorite agenda>. If we want to promote X with public money, do so explicitly and separately: support research, fund startups, etc. My 2c
Sorry for confusion. In "buying widgets" I meant the regular procurement decision: the DoD needs N units of something (door handles, tank shells, whatever), it publishes specs, asks for bids and buys from whoever can deliver at the lower cost to the taxpayer. The government also funds research, where it gives money to generate something it believes to be beneficial longer term.
Funding research is, by definition, less cut-and-dry as to what we should pay for; thus having an agenda is not always bad and might even be good. I am using an "agenda" not in a narrow political sense, but including positions like "we should be funding space comms / drone networks / real-time soldier health monitoring / whatever because industry is not building what we think we will need in a few years".
But being somewhat exposed to the waste of DoD procurement I am personally vehemently against inserting such ideology into procurement decisions. Those should be money-based. Get what you need at the least cost to the taxpayer. If you do not know what you need, think harder or invite experts or do a study (and publish it so people writing it know they are associated with those decisions) before paying billions for questionable junk. My 2c.
Eh, national security works differently from promoting healthy businesses.
The govt already has various programmes to help promote small business contractors in US defence. This is not a programme; it's a definitive project that has a specific set of (admittedly vague) objectives in mind. It's more efficient for the taxpayer for these to be accomplished when the funding is consolidated to a few entities for a 50% success rate, than to 20 different entities for a 5% success rate.
I’m all for a competitive commercial space type approach to gov projects and contracts but I really don’t see what that has to do with this.
DoD is experimenting with LLMs and is using multiple of the top providers in the space… just like every other tech company is doing. Everyone I know is coding with Claude, Gemini, or GPT and my experiments with Grok 4 have easily been as good.
If this was an innovation fund, ala what Canada likes to waste money on - where the gov pretends it’s a really bad VC, I’d at least understand these critiques.
Why would the DoD give them money just to break-even (not even likely. oai has easily put > 10m into their compute) with companies that already have the infrastructure? This isn't a stimulus package, it's a service they're buying.
Even then, these companies aren't doing research into LLMs, they're just wrapping the endpoints and creating some abstractions.
DOGE hates resellers and is trying to force all transactions through the prime contractors. Most small companies cannot afford to float the terms required by the federal government without resellers.
This admin is about graft and shakedowns. Just like the implosion of science, the companies that exist due to smallish federal contracts for obscure tech and speculative investments are toast.
> $10M could keep a startup running for 1+ years at its most crucial time. That's 10 solutions instead of 1 -- statistically one of them will be a massive breakthrough?
The failure rate for startups is much higher than 90%. And there’s the additional complexity of how do you pick which 20 such startups get the cash.
See my response to the other posters with the same notes
On the picking: it’s really not hard to search for AI companies and pick 20. In fact there are government programs that invest in startups so clearly it’s doable.
Who are those 20 companies? What would $10M do in the context of training LLMs that are competitive with Claude/O3/Gemini?
> That's 10 solutions instead of 1 -- statistically one of them will be a massive breakthrough?
The statistic is that 10% of startups make a massive breakthrough? Would love to see some work that comes remotely close to replicating that! Startup investing would be trivially easy.
Responded to the other poster with the same question.
Everyone says 1 out of 100 makes it big but the top 5-10% of a portfolio is still substantial. If we’re only giving the money to companies with revenue the odds of success are likely improved.
Startup investing is trivially easy. You give money to good companies and founders. There’s just a bunch of BS that gets in the way. Like giving massive money to big corps that don’t need it instead of startups that do.
Here are some top AI companies per Crunchbase: OpenAI, Anthropic, xAI, CoreWeave, Glean, Perlexity, PlayAI, Cohere, Tempus, Cyera, Replit, Windsurf, Mistral, Anysphere, Scale, Harvey, Thinking Machines Lab, helsing, Cluely, Suno, Clay, Crunchbase (lol), Lubega Geoffery, Caris LIfe Sciences, C3 AI, Runway, LangChain, Rigetti Computing, Cowbell, Laurel, SoundHound, Voxel, Harmonic, Builder, ElevenLabs, Decagon, Spring Health, Lovable.... alright I have a meeting to get to.
OpenAI - DoD invested in them and now I guess you agree with it!
Anthropic - same as above
xAI - same as above
CoreWeave - Doesn't make LLMs
Glean - Doesn't make LLMs (wow this startup investing thing might be harder than for you than you thought!)
Perplexity - Has finetuned LLama models AFAIK. Maybe you think Meta should've gotten the nod from DoD as well?
PlayAI - AFAIK only voices
Cohere - Not sure if they are LLama or otherwise
Cyera - Doesn't make LLMs
Replit - Doesn't make LLMs
Windsurf - Doesn't make LLMs
Mistral - Does make LLMs, you got one! Is French, though.
Anysphere - They make an IDE called Cursor
Scale - Doesn't make LLMs, basically a Meta subsidiary (you really must have wanted Meta to get the nod too!)
Harvey - Legal focus, not general
Thinking Machines - Mira Murati's company, just started 5mos ago, no public products. Definitely don't fit your definition of "has revenue"
helsing - Hadn't heard of them, are German.
Cluely - LOL
Suno - If the DoD gets into music generation this would be a great choice.
Clay - Don't know them, doubt they have LLMs.
Crunchbase - lol is correct
Lubega Geoffery - No idea
Caris LIfe Sciences - Life sciences doesn't sound right!
C3 AI - Scam
Runway - Media generation, not general use
LangChain - Doesn't make LLMs
Rigetti Computing - Dude, come on. They're a quantum computing company
Cowbell - Don't know them, but a google shows they're an insurance company lol
Almost all the rest don't even have anything to do with AI. So all-in-all, nearly a complete failure at suggesting even close to 20 alternatives for the DoD to invest in. Your answer didn't even hit US companies that do have some alternatives: Meta, MSFT, AMZN, SSI maybe?
I understand the sentiment to create healthy market but only a few handful company than can create general use LLM, most of them is just wrapper or small fine tune model for specific use case
I don't for anyone doing serious work. Use the Gemini family, O3 and Claude if you want to gsd. The DoD made the correct call IMO. Kimi K2 is also potentially interesting for non-defense purposes but I haven't spent enough time with it yet.
> My fund will never be $1B but that’s fine :)
It's "trivially easy" for you and 10% of your investments are expected to have "massive breakthroughs". Your strategy of "give money to good companies and founders" should easily enable you to reach $1B AUM!
not that they're great or very remarkable, but they can get the job done, and in a world where anthropic tries to assert dominance by fucking cursor seems relevant to the future to have your own models
As an American. I'd rather a single well established player get a large contract and actually deliver, than 20 disjointed companies each get 1/20th of the problem, have to work in concert, and possibly not even deliver at all.
This isn’t a grant to push for innovation. This is a promise from the orange man administration to the people and companies that donated to his "inauguration fund"
This is a kleptocracy but with extra steps. People are unfortunately numb to it.
