Tanuj Thapliyal

On Monday I said China is winning the AI race and we raised $12M to help fix that. 2M+ views later, the most common reply was some version of "what race? it's fear mongering. why not just slow down?" Because you can't! Open weight models are now roughly as good as the closed ones. You can download them today. The chips to run them are everywhere, in your laptop, in your phone, in racks all over the world. Slower, sure. But intelligence is something anyone can get their hands on now. That cat is not going back in the bag. Nobody has that power. Not Dario, not Elon, not the government. So slowing down doesn't stop AI. It just decides who has it. Slow down and the only people running frontier intelligence are sovereign governments and a handful of $3T companies. Everyone else gets locked out. That's the most dangerous version of this, and it's the one we walk into by accident if we hesitate. If it's here and it can help everyone, it needs to be available to everyone. Water, power, the towns that host these sites, and who actually pays for all of it are the right concerns. We just can't steamroll them in the name of speed. I've spent 15 years inside this physical layer, and getting the details right matter. Pay your own freight, build behind the meter, get audited. None of that requires going slow. It requires being disciplined while going fast. That's the whole bet behind Kos. Every American should win this, not just the ones who can afford to.

china isn't beating us on models..they’re beating us on invoices

If America loses the compute race to china it will be because of the speed of committed capital to steel in the ground.

Slowing down won't stop superintelligence. Here's what will actually happen if we slow down on AI: > Only sovereign governments and $3T companies will be able to afford and run intelligence > Everyone else will get locked out of this technology > We’ll see one of the biggest concentrations of power by another name > Everyday consumers get priced out of life saving medical tools and basic learning assistants Let’s compare AI to electricity. When did we start to feel “safe” about it? When we built so much of it that every town had a grid and every grid had an inspector. Look at the early sentiment for railroads. People protested the tracks because they thought steam engines would tear society apart. AI compute will follow the same legacy. Abundance will be the democratizer.

I spent 15 years building energy software for data centers. The public debate around data centers is stuck in a false binary between hitting pause and building blindly. Both sides ignore the one part that’s bleeding capital - Admin and back office work If AI is going to cure diseases and transform education, infrastructure needs to scale dramatically. That scale cannot be sloppy. Towns push back because data centers cannot freeload on local ratepayers for electricity and water. Operators must pay their own freight. The actual bottleneck is unverified contractor billings and un-audited megawatts. Every dollar lost to back-office waste keeps compute locked in an elite oligarchy. We need to track every billing deviation and audit the line items before payments leave the account. Grid pressure does not get solved by hitting pause. It gets solved by building faster with strict fiduciary discipline.

AI is going to get DIRT cheap We might even reach a point where it takes 0 water to use AI I know how that sounds I’m not trying to ragebait anyone so let me explain There is a part in the AI cost curve that’s being hidden from you Everyone talks about Nvidia pricing power and next gen chips. They want to make you believe that AI compute will keep getting more expensive because demand is exploding. I can (with full confidence) say that this is 100% false How do I know? I’ve been in this field for 15 years now I started by doing AI research at stanford The thing that people end up forgetting is that hardware is one component of the bill On a $1B to $2B data center build with 8,000 workers on site, up to 30-40% of the total cost is administrative and contractor bleed. 400-page PDFs….. contractor pay applications….3,000 line items per invoice….Equipment rental markups…. These are the single LARGEST hidden cost in the entire AI infrastructure buildout. The waste is systematic. Cheap ubiquitous compute is a financial controls problem We’re fixing this now and when it’s done AI will be dirt cheap (and safe)

Americans are struggling everyday… Electric bills are doubling because data centers are tapping straight into local power grids. Towns are RIOTING over server farms drinking municipal water while transmission lines hit capacity. But let me just make this claim right now AI is NOT the problem here City commissioners think blocking new facilities protects their residents. That’s the problem Blocking domestic data centers only makes compute scarce. When compute gets scarce, only the richest people and largest companies can afford it. That builds an oligarchy. Environmental responsibility has to stay front and center. But the solution is not slowing down If a data center director jacks up local utility rates or drains local water New permits have to factor that in. Tech firms need to build their own power behind the meter and pay their fair share of taxes This stretch of time we’re going through right is very unique We are literally in active labor right now. We’re birthing the age of AI Contractions hurt and labor is bloody, but the baby arrives regardless. America either builds with discipline or loses to China.

Everyone on X is lying to you They say “China is going to win the AI race” Yes they will…But the reason is not what you think Everyone on the internet believes it’s environmental pushback and energy shortage That we need to “slow down” I think we need to speed up (now more than ever) Don’t get me wrong Protecting local grids and water is non-negotiable, being a responsible fiduciary is non-negotiable But If compute stays scarce, only the rich get to use it. Ubiquitous, cheap compute is the only way America beats China and makes intelligence accessible to everyone. The reality is of AI is not models and AGI AI is concrete, steel, chips, equipment, power plants, substations, and transmission lines. That’s where we’re losing We’re wasting money and infrastructure on things that don’t need it. We’re so excited to discuss the “$50B AI deal” that happened last week But what happens after? That capital most likely ended up in an analog back office. With a bunch of contractors reviewing that line by line Billions of dollars bleed out through these contractors And THAT is why AI is expensive and slow Trust me, I spent the last 15 years inside that physical layer. I started in Stanford AI research under Sachin Katti. I built networking hardware for server racks, wrote energy efficiency software for data centers, and deployed computer vision on oil rigs and construction sites at Spot AI. I watched the money pipeline for years This contractor chaos and pure wastage are the reasons compute stays artificially scarce and expensive. This is what i’m building now A financial layer so strong that compute becomes cheap and easily accessible Because if we want cheap AI for the world, we have to fix the back office first. Lots of opinions, exciting things and announcements coming soon Follow along if you like nerding out on AI, Tech and Financial systems

800-page invoice. Reviewed in seconds. By an employee that doesn't exist. Meet @kos_ai_inc . The world's first virtual finance employee. Own computer. Own software seats. Own identity in the org chart. You talk to it over email, Teams, Slack, Zoom. Teach it new workflows over screenshare like a new hire. Datacenters. Energy. Defense. Factories. @8VC @xyz_vc @MVPVentures @kos_ai_inc kos.ai