Building apps @honeyflowhq, @sweptdating, and running @talegeninc | AI guy | ex-Amazon | Avise La Fin | 👉 Comments are my own
United States
Joined February 2013
- Tweets13.6K
- Following912
- Followers1.6K
- Likes6.5K
OpenAI plans to buy Google's chips. The Information reported it Tuesday. The company that made Nvidia indispensable is now adding a second supplier, after renting Google's TPUs since June 2025. The moat was never just the chip. It is who can supply cheap compute at scale.
Thirty minutes. The decision lands at 2 PM, Warsh talks at 2:30, no dot plot this time. 94 of 100 economists polled by Reuters expect a hold at 3.75 to 4.00 percent. All 100 expect a hike in December. The buildout borrows at whatever number comes out of that room.
The bottleneck moved. It is no longer the chip. It is the connection and the permit, and the states writing moratoria are deciding who gets either one first.
And the money is not leaving power. Crusoe just closed a $700 million facility secured by the power generation equipment itself. Gas generators, behind the meter, as collateral. When the asset is the generator, the lender does not have to bet on the lease.
The watchdog count for Q2 alone: at least 45 projects worth roughly $68 billion blocked or delayed. That is not a bubble popping. That is a permitting system saying no, one docket at a time.
Only New York has actually passed a statewide pause so far, per the state-legislature count. The rest are bills, proposals, and fights at the statehouse. A bill is not a ban, but 16 of them at once is a signal developers read before they break ground.
What good is $765 billion if 16 states can pause the building? Goldman counts $1 trillion of AI infrastructure spending next year. The statehouses are counting moratoria bills. Somebody has to win that argument before the first shovel moves.
Germany's grid queue is 147 gigawatts long, and datacenters are waiting behind batteries. Berlin now plans to legally protect grid access for datacenters and industry. When the queue is the policy, allocation is the whole fight.
Dating apps all have the same fake-profile problem. Swept is being built the other way around: prove the person is real first, then worry about the match.
Altman says the next loop is robots building robots and datacenters building datacenters. A loop like that does not bottleneck on labor. It bottlenecks on power and permits. The countries that issue both fastest get the loop.
teapawt.org/rohanpaul_ai/status/21…
Robots building robots is the next big loop
Asked what infrastructure OpenAI is betting on now, Sam Altman says its robots that build more robots, and data centers that help build more data centers, chip factories and power plants, so compute can grow far faster.
says this loop is almost as big a deal as RSI, AI improving its own research.
----
Full video on "Stanford Graduate School of Business" YouTube channel, (link in comment)
CNBC reports TikTok's US business is shopping for multiple gigawatts of compute, with one source putting the target at 4 GW by the early 2030s. A video app is now in the power market. That tells you the AI buildout stopped being a chip story a while ago.
My process at Talegen is boring on purpose. Discover: map the workflow. Prioritize: find the one handoff costing the most time. Build that. Improve after it proves value. No AI strategy decks. A working system.
Generic CRMs fail direct sellers. Their business runs on parties, reorders, and remembering what a customer bought 6 months ago. HoneyFlow is designed around how a Mary Kay consultant's week actually works, not how Salesforce thinks it should.
"Data centers don't create jobs" is the laziest line in the permitting fight.
An 80 MW facility: 125 jobs inside the building, 800 more across 14 industries around it. 925 total, a 7.4x multiplier. A 500 MW build runs $9.5 billion in, $1.5 billion a year out, $1.52 billion in tax revenue over ten years.
The building looks empty from the road. The paychecks aren't.
The oversupply take has a second life after CNBC. It becomes the exhibit in every permit hearing from here on. 'Even Gartner says we have too much.' The analysts are writing the blockers' talking points for them.
teapawt.org/JonahLupton/status/211…
Hmmmm... how can this data/chart (see below) be correct... after the bozo from @Gartner_inc went on CBNC and said we already have too much supply from ai data centers... despite having ZERO tech executives, industry experts, ai startup founders, investors and sell side analysts confirming this nonsense.
I have not seen one data point to suggest that supply is ahead of demand... and one could aruge that demand is accelerating with open weight models and cheaper tokens which still require more ai infrastructure to be built.
Gartner is just trying to stay relevant... putting out flawed analysis then doing a media tour to spread the FUD is not a great strategy... they lost a lot of credibility this week.
I suspect they'll be back to irrelevance in the next few weeks as we get Q3 earnings when megacaps / hyperscalers will confirm they're still compute constrained and need to keep adding capacity to meet increasing demand.
Tonight we get $SPCX earnings and tomorrow is $META earnings — if they do give 2027 capex guidance I’m expecting at least $400 billion (combined)… and for $META that doesn’t even include their neocloud deals.
If @elonmusk and @SpaceX are going to add another 6+ GW of capacity before the end of 2027… they’ll need to spend $240-300B in capex over next ~15 months.
Currently the sell side is looking for $META to spend $200B in 2027 capex however those estimates pre-date the launches of Muse and Wonder so I suspect $META's 2027 capex number will be $250B+ and could be closer to $300B unless they expand relationships with neoclouds ie $NBIS $CRWV and Crusoe.
Meta reports after the close. Expected: $41B in Q3 capex, negative $9B free cash flow, $63B in revenue. The stock is up 37% since July, so the spending is priced in. Tonight's trade is one line of the call: the 2027 capex number.
Mary Kay gives consultants InTouch. Treat it like a ledger: it's where you reconcile. HoneyFlow is where the actual work happens. Bookings, follow-up, events, inventory, in one flow. That's the product I'm building.
The umbrella is the whole trick. Streets make the footage, courts make the delay, city hall makes the access. That is a division of labor, not a coincidence.
teapawt.org/DataRepublican/status/…
