Jeremy Kent

[Jeremy Kent @jdk3154]

Arguably the most detailed compute disclosure produced landed Monday, and it still can't tell you whether we got more efficient.

Anthropic S1's compute line: 201% of revenue in H1 2025, 55% in H1 2026. Pricing, amortisation and efficiency all land inside one line.

Huh.

[Jeremy Kent @jdk3154]

As a former consumer goods product manager and Apple fan boy, this one hurts.

https://t.co/pjtUU1SJhb

[Jeremy Kent @jdk3154]

Fascinating read by @ttunguz on inference economics in a fast changing landscape.

The name of the game is increasingly becoming compute utilization.

[Tomasz Tunguz @ttunguz]

Yesterday brought two data points about the health of inference businesses : SpaceX's third-quarter report & DeepSeek's inference post.

Overall, they demonstrate that inference businesses can be quite healthy even at current prices & efficiencies. https://t.co/cum1Vbah9Z

[Jeremy Kent @jdk3154]

"Stranded compute" might just be the catch phrase of 2027.

Hard job for CFO's these days to measure ROI / risk vs return on traditional vs. neoclouds. I'd be looking at software-centric solves (there are more and more energing).

[Chamath Palihapitiya @chamath]

As someone with a portfolio of energized land assets here is what I am seeing:

1. Neoclouds are disorganized and poorly capitalized. Some are definitely better than others however.

2. Hyperscalers are intractable, bureaucratic organizations that take 3 months to do an NDA and want 18mo escrow periods. We built our portfolio by competing against these terms because of our speed and simplicity. The more future assets orgs like ours win, the bigger the problem for the hyperscalers downstream (see 3).

3. There are delays in the existing pipeline that will catch people off guard in 2027. Hyperacalers will (obviously) react by focusing on their own internal compute needs over any third party. If the third party is a model company, they will tell them to go through their model router where they will also take a non-trivial vig. Them telling Ant they have no compute for them (rumored in Ant’s S-1) if true is a huge tell that this is already happening.

All of this is to say that it does look like there will be a lot of partially stranded compute and compute sitting in inventory in 2027 than is currently forecasted. This is bought and paid for but effectively underearning (below total cost of capital) or collecting dust. We see this as true for everyone - model companies, hyperscalers, neoclouds except ONE (you can guess which one).

This dislocation will benefit two kinds of orgs: the non-bureaucratic big scalers who can build industrial scale quickly and the nimble, well capitalized players feeding them.

2027 will be a turbulent year because it will clear out a bunch of wrong/bad bets. It will also be a consolidation year for a few.

[Jeremy Kent @jdk3154]

@SemiAnalysis_ Very novel approach and arguably the most complete public accounting of supply that exists.

[Jeremy Kent @jdk3154]

Meta's Q3 call. Zuckerberg on compute: "one of the limiting factors over the next year."
CFO Susan Li on the near-term lever: "off-peak or spare computing capacity."

The largest buyer of accelerators can't even find a shop to buy more compute next year.

[Jeremy Kent @jdk3154]

Big news in regional business. Huge congrats to Roland, Lisa and the entire OnLogic team for this strategic new era.

https://t.co/aaeiimjEGo

[Jeremy Kent @jdk3154]

@a16z Super interesting take.

[Jeremy Kent @jdk3154]

Its staggering in an age of compute scarcity how much of the daily compute is work that has already been done on a GPU. This new .ONCE is a very slick way to save a ton of cost and time.

[TAHO Labs @TAHOlabs]

Check out this video of .ONCE, one of the ways TAHO gets you more from the GPUs you already own.

Same prompt, same machines, sent twice. The second comes back complete immediately: the work was done once and reused.

[Jeremy Kent @jdk3154]

Looks like @gnodis over at @TAHOlabs called this one early. Here are the receipts:

https://t.co/TD9OgDErqt

[Nathan LeClaire @dotpem]

I can't emphasize enough how important this @a16z chart seems and how few have internalized it.

[TAHO Labs @TAHOlabs]

Two weeks of memes, one idea underneath all of them: the compute you're looking for is already in your rack.

The detailed version is 20+ essays from the people in the lab building TAHO. Start anywhere.

https://t.co/RQElgDHusj

[TAHO Labs @TAHOlabs]

TAHO Labs SOC 2 Type 1, completed September 2nd.

TAHO Labs runs beneath your stack on hardware you already own — your model, your scheduler and your data path don't move.

Controls and subprocessors in the Trust Center.
https://t.co/OeZmAyafNi

[Todd Smith @dynadog44]

Things that never show up on the GPU bill:

1. Test machines nobody ever turned off.
2. Jobs that quietly failed and ran again.
3. Software sitting in a loop asking if the answer is ready yet.

When considering that you pay people to fix it, the bill is the cheap part.

[Jeremy Kent @jdk3154]

There are two ways to have more compute next quarter.

1. Buy it. If you can get access to it (which is getting harder even for bigger players).

2. Convert more compute out of what is already owned & installed. That is already funded and is becoming easier by the day.

[Jeremy Kent @jdk3154]

A GPU fleet can be 97.9% rented and still convert only 15% of its throughput into work.

The first figure got press this week from Oracle Q1 reporting. But second figure - capacity used on a chip, is where the real story lies.

[TAHO Labs @TAHOlabs]

You can't add a machine to a job that's already running. Everyone who's run infrastructure knows that...until now.

Here's three machines halfway through an AI job. We plug in a fourth, it picks up work already in progress, nothing restarts, and the answer finishes with its help.

[Jeremy Kent @jdk3154]

Every conversation about compute efficiency measures the prize and skips the real cost: what has to change, what breaks, who has to be in the room, how long
the dual carrying cost runs.

That second number set is why efficiency deals die after the meeting goes well.

[Jeremy Kent @jdk3154]

Every AI compute efficiency pitch in this growing market asks you to move something. A migration, a new endpoint, a different model, another vendor in the data path.

The ideal version of the problem is the one where the capacity is already in your building & nothing has to move.

[Jeremy Kent @jdk3154]

Great day at the Exit 5 DRIVE conference (B2B Tech Marketing) in the green mountains of Vermont. Some bleeding edge GTM Eng insights amidst beautiful scenery.

[Jeremy Kent @jdk3154]

Meta trained Llama 3 on a purpose-built 16,000-GPU cluster, with one of the strongest infrastructure teams in the world, and peaked in the high 30s to low
40s percent chip utilization.

That's the ceiling under close to ideal conditions. Begs the question...