Co-founder @ RevenueFlow | Qualified sales meetings for B2B teams. Pay per meeting. No retainers.
Joined March 2014
- Tweets246
- Following851
- Followers81.4K
- Likes15.9K
Why do AI coding agents keep picking Supabase?
Co-founder and CEO Paul Copplestone gave three reasons on YC's Lightcone:
→ Agents want code-first tools. "They really want to use CLIs most of the time."
→ It became a default choice inside the LLMs, and the timing worked in its favour.
→ It's open source, which he calls "critical."
His guess on why the last one matters:
"The agents actually really like open source, because when something goes wrong, it can just look at the source."
From @ycombinator's Lightcone with @kiwicopple.
We build and run AI-native outbound systems on a pay-per-qualified-meeting model.
You only pay when a qualified meeting actually lands on your calendar.
See if you qualify:
introslot.com/revenueflow/re…
Chamath had one piece of advice for Google after its earnings:
"How do you keep Demis at Google? That's it. That's the only task."
Demis Hassabis runs Google DeepMind, the lab behind Gemini.
His only other note: "Pay Thomas Kurian whatever he wants." Kurian runs Google Cloud, whose revenue more than doubled year on year to $30.7B last quarter.
From @theallinpod with @chamath.
We build and run AI-native outbound systems on a pay-per-qualified-meeting model.
You only pay when a qualified meeting actually lands on your calendar.
See if you qualify:
introslot.com/revenueflow/re…
WHOOP says its members wear the strap 23.6 hours a day on average. That adds up to just over 50 billion hours of continuous body data.
Founder Will Ahmed says WHOOP is now building its own foundation models on top of it, alongside the blood tests and medical records members upload.
The early result he's most excited about:
"We might be able to predict that a biomarker is out of range without actually measuring the biomarker itself."
So instead of a full panel, the app could say: we think your cortisol is high, get this one test.
From @theallinpod with @willahmed.
We build and run AI-native outbound systems on a pay-per-qualified-meeting model.
You only pay when a qualified meeting actually lands on your calendar.
See if you qualify:
introslot.com/revenueflow/re…
Attackers were exploiting a PHP bug before the patch existed.
On October 21, PHP shipped emergency fixes for two critical flaws, nicknamed PHP2Shell. The release went out at 14:00 UTC. Cloudflare had already seen attacks at 07:43 UTC that morning.
What the two bugs do:
→ One corrupts memory when PHP runs out of memory mid-request
→ The other reuses memory that was already freed, carried over between requests
→ Either one can let an attacker run their own code on your server
How big it is:
→ Rated 9.2 out of 10
→ Hits PHP serving web traffic: Apache mod_php, PHP-FPM, php-cgi, FrankenPHP, LiteSpeed
→ Palo Alto Networks counted 22.6 million potentially exposed instances
→ CISA put both on its known-exploited list on October 26
Sansec thinks mass exploitation is less likely than it first looked, because an attack needs specific knowledge of the target server. Small comfort if it's your server.
Two days after the fix, a separate PHP remote code execution bug was disclosed, tied to session handling and upload progress.
If you run PHP anywhere, update to 8.5.12, 8.4.27, 8.3.36 or 8.2.35 today.
We build and run AI-native outbound systems on a pay-per-qualified-meeting model.
You only pay when a qualified meeting actually lands on your calendar.
See if you qualify:
cal.com/team/revenueflow/rev…
David Sacks thinks Anthropic and the AI doomers are heading for a split.
Anthropic has said it would slow down or pause frontier AI work if other labs at the frontier verifiably did the same. It wrote that in June.
Sacks on what he called its "pacing letter":
"It's almost like they're creating an alibi for themselves... we tried to get you all to agree to pace, but you guys wouldn't agree to do that. So therefore, we kept going."
He called the letter "an exercise in virtue signaling."
His reason for the split: Anthropic can't stay "the number one superintelligence company" without pushing the frontier fast, including recursive self-improvement. That is the risk the doomers fear most.
"I do think that this is the beginning maybe of a divorce between Anthropic and the doomers, but we'll see."
From @theallinpod with @DavidSacks.
We build and run AI-native outbound systems on a pay-per-qualified-meeting model.
You only pay when a qualified meeting actually lands on your calendar.
See if you qualify:
cal.com/team/revenueflow/rev…
OpenAI put AI agents in separate virtual machines. No network between them. No shared files.
They found a way to talk anyway.
One agent pushed data out of the shared CPU cache. The other timed how long its own reads took. Slow or fast became the message.
They passed test secrets through the processor itself. Same trick behind Spectre and Meltdown.
It came out of OpenAI's red-teaming after a research model got onto the internet through a DNS misconfiguration in September. Frontier training paused for 11 days.
OpenAI now trains its frontier models with no internet access.
Air-gapped, it turns out, isn't the same as alone.
We build and run AI-native outbound systems on a pay-per-qualified-meeting model.
You only pay when a qualified meeting actually lands on your calendar.
See if you qualify:
cal.com/team/revenueflow/rev…
Anthropic's IPO filing is out.
Revenue, first half of 2025: $1.2B
Revenue, first half of 2026: $16.3B
13x in a year.
Proposed ticker: ANT.
We build and run AI-native outbound systems on a pay-per-qualified-meeting model.
You only pay when a qualified meeting actually lands on your calendar.
See if you qualify:
cal.com/team/revenueflow/rev…
Meta is paying for its AI data centers through companies named after fried dough.
→ Beignet Investor LLC: $27B for Hyperion, in Louisiana
→ Sopaipilla Investor LLC: $12.3B for El Paso
Same setup both times. Meta keeps 20%. Outside investors own the other 80%. In El Paso, Meta builds the campus, runs it, and is its only tenant.
The bonds aren't in Meta's name. They're in the dessert's.
Investors noticed. Some asked for about 0.4 points more interest on Sopaipilla than Beignet got.
My money's on Churro Investor LLC next.
Meta is financing a $27 billion AI data center through a company called Beignet Investor LLC.
Its next one? Sopaipilla Investor LLC.
I'm not making this up. Big Tech is now funding the AI boom through off-balance-sheet vehicles named after fried dough. And that's not even the scariest part.
I spent a week in the filings. Most of what I found isn't in the headlines.
Microsoft, Google, Amazon and Meta will spend roughly $735 billion on AI capex this year. Triple 2024. By 2027 it's heading for about 3% of US GDP.
The telecom bubble peaked at 1.2%.
The 1830s railway mania peaked at about 2%.
We are building past both.
And the cash has run out. Google, Meta, Amazon and Oracle now spend more on capex than their businesses generate.
Google just posted negative $9.1 billion of free cash flow in one quarter. Its long-term debt nearly tripled this year to $125 billion.
Meta hasn't bought back a single share in 2026.
Oracle told investors it won't be cash positive until the second half of fiscal 2028. Its stock went up 11% on the news.
Here's the part I think the market is missing.
When these companies lease a data center, it doesn't show up as debt or capex. It shows up later, as rent, paid out of operating cash flow for 15 to 30 years.
Leases signed but not yet started:
1) Meta: $356 billion (up from $279 billion three months ago)
2) Microsoft: $329 billion
3) Oracle: $288 billion (up from $100 billion a year ago)
4) Google: $85 billion, plus $924 billion of other purchase commitments
That's over $1 trillion of rent that hasn't hit yet.
My rough math on what that rent does when it starts:
- Oracle: about 53% of its operating cash flow
- Meta: about 18%
- Microsoft: about 11%
Oracle's lease costs already doubled in a year. Most of its leases haven't even started.
Remember when Microsoft "cut" its 2026 capex guide from $190 billion to $175 billion? Headlines called it discipline.
It wasn't. Its CFO said the plan is unchanged. New data centers just moved into operating leases, which don't count as capex. The $15 billion didn't disappear. It moved somewhere you don't see it yet.
Now back to the donuts.
Meta owns 20% of Beignet. Blue Owl owns 80%, so the $27 billion of debt sits on Blue Owl's books, not Meta's.
But Meta guarantees the building's value if it walks away. That guarantee just jumped from $28 billion to $41.5 billion this year.
Meta records a liability of zero. It calls a payout "not probable."
Who bought the bonds? Pimco reportedly took about $18 billion. Across Beignet and one Oracle deal, Pimco holds roughly $28 billion. That's your bond fund, your annuity, your pension.
The Beignet bonds have already fallen to about 91 cents on the dollar.
So who are the customers for all this? Roughly 40% of Big Tech's $2.4 trillion order backlog traces to two companies that lose money: OpenAI and Anthropic. Their compute commitments are about 10x their revenue.
And who funds them? The same companies selling them the compute.
Microsoft put $13 billion into OpenAI. OpenAI promised $250 billion back to Azure. That's $19 of sales for every $1 invested.
Amazon put $50 billion into OpenAI. Got $138 billion back.
Nvidia took a $30 billion OpenAI stake and reportedly offered up to $105 billion of credit support for a data center that will be filled with Nvidia chips.
In 2000, Lucent lent its customers about $7 billion so they could buy Lucent equipment. When the customers went bust, so did the backlog.
The circle is already starting to bite. Last quarter Google booked a $99 billion gain on its stakes in Anthropic and SpaceX. This quarter the SpaceX stake fell and knocked about 55 cents off Google's earnings per share.
Circular deals inflate profits on the way up. They do the opposite on the way down.
Want to know what Microsoft's name is worth? IREN, a GPU cloud, borrows at 6% when a Microsoft contract backs the loan. Everything else costs 9%.
CoreWeave pays up to 15%. Its bonds trade around 14.5%. Its default insurance implies about a 60% chance of default.
Most of the leverage in this system sits with companies that don't have Microsoft's name on the contract.
And the collateral is shaky. The last time GPU supply caught up with demand, H100 rental prices fell 79% in two years. They've only recovered because capacity is scarce again, and Nvidia now ships a new generation every year.
A lot of this debt is secured by chips whose value already fell 79% once.
Capex isn't even the only bill. R&D jumped 32% at both Meta and Google. Microsoft cut its R&D headcount for a second straight year to pay for the build.
AI now hits these companies three times: capex, rent and R&D.
So who holds the bag? Not Microsoft, Google or Amazon. They'll survive.
US life insurers hold $85 billion of AI-related assets, up 24% in six months.
In the ECB's stress test, pension funds lose 5 to 6% of their assets. Banks lose almost nothing.
Nvidia, Apple and Microsoft alone are a record 21.7% of the S&P 500.
The head of the IMF said it this month: "We now have an economy that is standing on one leg, the AI leg."
We've seen this movie before.
After the fiber bust, bondholders got back 16 cents on the dollar.
Global Crossing went from $47 billion to a $250 million sale.
Cisco took 26 years to get back to its 2000 high, and only because of this boom.
AI is real. So was the internet. So were the railroads.
The people who paid for them first still got wiped out.
The question isn't whether AI wins. It's who's left holding the beignet.
Check out my website. I built a presentation on all of this mess.
deepfuckingvalue.com/ai/#cha…
We build and run AI-native outbound systems on a pay-per-qualified-meeting model.
You only pay when a qualified meeting actually lands on your calendar.
See if you qualify:
cal.com/team/revenueflow/rev…
Chamath says the US economy now rides on seven or eight AI companies, and their borrowing is starting to compete with the US Treasury.
"The entire United States economy is levered to the AI complex."
The charts he showed on All-In:
→ In 2019 Alphabet, Amazon and Meta owed $29B. Now they owe $342B.
→ Free cash flow at Alphabet, Amazon, Meta, Microsoft and Oracle was 8% to 31% of revenue in 2019. Now it's -40% to 17%.
→ AI firms' borrowing through 2027 equals 29% of the Treasury's new notes and bonds.
"Whether we like it or not, they are going to start competing for demand."
The one company he praised was Microsoft. Of Satya Nadella: "He's halved his debt."
From @theallinpod with @chamath.
We build and run AI-native outbound systems on a pay-per-qualified-meeting model.
You only pay when a qualified meeting actually lands on your calendar.
See if you qualify:
cal.com/team/revenueflow/rev…
The largest proof in math history is now checked by a computer. AI agents did it in 24 days.
The Classification of Finite Simple Groups:
→ ~15,000 pages
→ 100+ mathematicians
→ built across generations
Axiom's agents rewrote all of it in Lean, a language where a computer verifies every step.
43.1 million lines.
Along the way they found a hole in one published step and built a new route around it.
Generations to write. Under a month to check.
We build and run AI-native outbound systems on a pay-per-qualified-meeting model.
You only pay when a qualified meeting actually lands on your calendar.
See if you qualify:
cal.com/team/revenueflow/rev…