GP @20VCfund. Obsessed with investing. 2x Father. Investor in: @SolveIntel @ClickHouseDB @FireworksAI_HQ @FuseEnergy @Rivanindustries @airwallex and more
London, England
- Tweets 1,492
- Following 786
- Followers 16,620
- Likes 132,757
I like this type of corporate communication.
When you mess up, just say so and apologize. Then try hard not do it again. Works with everything in life.
Opus 5 was indeed terrible to use. I nearly churned. But now I can’t wait to update all my coworks to opus 5.5!
3
2
20
3,555
This is the worst thing that can happen to Neolabs, by the way.
Neolabs who told VCs they needed $500m for compute in March / April may now only get 50% as far, if they haven’t committed to fixed long term contracts yet.
So, who will put in the extra hundreds of millions?
Ornn's B200 Rental Index has reached an all time high.
B200s on-demand rentals are now transacting at an average price of $7.88 per GPU hour.
data.ornn.com
10
3
65
26,199
30 years later: are we living Amazon’s own Barnes & Noble moment?
1996: Len Riggio takes Bezos to dinner, tells him B&N's website will crush Amazon, then floats a partnership. Bezos passes
2026: Amazon passes on Muse. Bets its own agent owns the next interface
Interesting.
10
3
27
5,818
The best consumer value proposition wins.
Amazon, of all businesses, should know that.
If customers wants AI agents to shop for them and Amazon won't sell to agents, someone else will.
There will be generational companies built on creating the ultimate experience for agents.
5
1
14
1,381
Feeling blessed my wife did not expect me to be post-economic when we met 7 years ago.
Two baby girls later (and more to come!), I am grateful for the bet she made.
Want a great marriage?
Think first about what you can give. Because it’s a team sport.
31
8
542
63,045
Extremely smart move by @FactoryAI.
On-prem / own VPCs may become a much larger part of workloads with AI, as large enterprises will care more about protecting sensitive information and controlling costs.
Introducing Factory Private, a new deployment option that lets organizations run Factory in their own VPCs, on-premises environments, or air-gapped networks.
2
4
32
11,135
No…
Snapchat will *never* truly crush it until it stops being a ponzi scheme that puts shareholders money into employees pockets via ridiculously excessive share based comp (for no results).
I think @evanspiegel is actually making sick progress on SPECS and Snap is going to crush it when these get to < $400
4
18
6,741
Frontier AI won't slow down.
Nobody slows down unilaterally… and coordinating is illegal without an antitrust waiver.
However, let's say it did. What are the implications? 💭👇
*1) Does it reduce demand for AI?*
In the short to medium term: no.
Most companies are still wrestling to figure out how to implement it, with a great ROI. The frontier has run much faster than the use cases so far.
You can solve extremely complex math problems with the best models - yet we are only just starting to have useful AI assistants.
Beyond consumer search, coding and perhaps law, penetration is still low.
Models are good enough to process insurance claims, customer service calls, and dozens of other use cases.
These do not need a better model. However, they need cheaper tokens and someone to build the workflow.
If anything a pause is great for the app layer. There would be less talk of "the next model eats my startup" - and the labs would need to start competing on price and distribution instead of capability.
And of course, truly useful consumer AI will be a massive source of demand too (muse, instinct, town etc…)
*2) Do we need any less compute?*
I spoke to a team trying to secure 1k GB200s on Friday… well even with tens of millions in hand, it's not easy. Demand vastly outpaces supply already.
Pacing, in Dario's own words, does not mean stopping training. It means slowing capability gains.
The compute keeps running. But it will move: more inference, more reinforcement learning, more world models and robotics.
Great occasion for adjacent labs to catch-up / overtake too. And given 1), inference demand compounds regardless.
So no.
*3) Will open source overtake the frontier?*
Well, that is the most interesting one. If open source truly innovates (instead of distilling and copying), then it should.
But open source today mostly means Chinese labs. And every big open leap has landed months after the equivalent frontier leap. Never before.
DeepSeek has real innovation on the training side (MLA, FP8). But a post-training gap closing so fast is hard to explain without distillation, and Dario clearly agrees (and spoke about this many times): his essay pairs pacing with a crackdown on distillation and weight theft.
So: open weights become the floor in terms of speed. The frontier can only pace as much as China lets it: the stakes are too high to fall behind.
*Conclusion: if it did, I would remain bullish apps and inference, but turn more bearish lab pricing power. Which is probably why it won't happen.*
7
2
29
2,449
Bending Spoons acquires Miro for ~$1.79b. But who gets the 💰? My usual breakdown below 👇
*This is 89.8% below the last round, and <3x EV/ARR* 📉
Yet, investors are making $370m of profits on a company that was once valued at $17.5b! Not a glorious outcome but a 1.78x blended multiple on $473m invested. 👀
There are mainly three parties doing well here: (1) founders & team, (2) @altairvc, and (3) @Accel.
Most of the capital came in at the top, and the preference sheltered investors nicely from a 90% haircut... Not a great return after 5 years of equity cost of capital, but "better 1.0x than 0.1x".
So let's dive in:
*1) Founders, Team and Common shareholders will be sharing nearly $1b* 🥳
While I am sure they expected much better in 2021 at the height of the bubble, this is an acceptable outcome.
Miro can thank its frugality for this outcome: (1) the acquisition press release implies it still has $435m on balance sheet at closing, and (2) the company only ever raised $473m!
The cash added back to the $1.355b enterprise value effectively paid by Bending Spoons really helps (<3x ARR), and only the Series C claims pref.
*2) Accel is getting the most $ back, but not without regrets* 🫠
They are making c.9x from the Series A.
But they will have mixed feelings today, and will regret following on in the Series B and C which will return ~1.9x and 1.0x respectively. Both of which are obviously not great for six and five years of equity risk, respectively.
*3) AltaIR likely still returns >100x ‼️*: investing at pre-seed can pay off, even when growth investors do not make money.
*4) Disappointing outcome for growth investors. But the pref worked as intended.* 😮💨
The acquisition is a 90% haircut to the Series C price of $17.5b. However, the Series C investors will still make 1.0x back thanks to liq pref.
ICONIQ, Salesforce, Atlassian, Dragoneer, TCV, DST and @codorniou can all thank this mechanism, without which there would have been a few very awkward LP calls!
32
33
388
70,544
@jasonlk / @rodriscoll some materials for the next Trio show ❤️
1
11
2,666
Been very long i have not been enthused by an Apple product. I am not more enthused today.
To start with, just make Siri actually intelligent / helpful.
And improve voice recognition from airpods on the move (pre-requisite for the above).
6
29
3,593
Updated tracker for OpenRouter four-week token growth 👇
Now at ~14% WoW. At that pace volume doubles every 5-6 weeks. Growth is accelerating: trend went from 2% WoW to 11% over the period.
This comes at a time when OpenAI cannot handle its direct traffic growth. Unprecedented
5
30
3,303
It was good fun discussing my article "The Great Neolab Trade" with the @etnshow boys, @lukeknight and @Ronanchamberss.
In fact, when I said "not many" - there are only four tech co's that cleared >$100bn val at exit so far.
The Neolab bet is even less obvious than you think.
20VC General Partner @PaulBonnet says that VCs investing in neolabs at $5B pre-seeds need to see a $100B return.
"When you invest at 5 billion seed or pre-seed, what you're really saying is that company will be worth at least 100 billion."
"When that bet is placed 10, 20 times over, you have to start asking, 'Where is all that value coming from?'
"I'm really hoping that we have some incredible breakouts from London but you [then] have to assume that that there are two more Anthropic and OpenAIs."
6
4
34
8,255
How much is $ORCL worth if the market now starts fully valuing the RPO from OpenAI?
Given Astra I would not be surprised to see a continued revenue acceleration.
I will be updating my public ARR tracker as soon as we have more news.
Meantime, I’ll take the bet it’s worth more.
3
1
16
3,355