Ismaeel Bashir

Increasing GPU utilization and effective capacity @ Expanse

Co-Founder/CEO of Expanse

[Ismaeel Bashir @ismaeel_bashir_]

@southpolesteve https://t.co/VxTINvTSGp

[Ismaeel Bashir @ismaeel_bashir_]

@southpolesteve I made a version of this back in the day: https://t.co/6xLHogkxle

as a joke - its a lot slower :(

[Ismaeel Bashir @ismaeel_bashir_]

We’ve raised a $5.3M seed for @ExpanseCompute, led by @crane_vc with PXN Ventures.

Expanse predicts exactly what an AI workload needs before it runs, so teams get as much as twice the work from the same hardware.

Why we started it: https://t.co/jBFSCTR0Bb

[AIP Seed @AIPSeedVC]

🚀 @AIPSeedVC invests in @ExpanseCompute 🚀

We joined the $5.3M seed round led by @crane_vc, with PXN Ventures and angel investors from DeepMinds.

Congrats and welcome to @ismaeel_bashir_ @nkdem_b and the entire Expanse team 🫶

[Ismaeel Bashir @ismaeel_bashir_]

Most cluster monitoring tools can’t tell you, within your allocations, whats being used properly. It tells you a node is allocated. It doesn’t tell you what’s happening inside the allocation. So a job that asked for 400GB and used 60GB reports as busy and you see ‘high utilisation’, but in reality there is 340GB sitting there doing nothing.

Across a cluster this reduces the number of jobs able to flow through your cluster. People then complain they dont have enough capacity, so you buy more, it takes 6-12 months and you still have the same problem.

We’ve measured this problem across various production clusters. Checking every job, what it asked for against what it went on to use. A lot of them reserved 2-3x what they used.

On one cluster the gap came to $8M of capacity nobody could see in a single month.

I’m talking about this and more at @stacresearch in New York in October, drop me a message if you’ll be attending :) - we will also be at the London event as well.

[Ismaeel Bashir @ismaeel_bashir_]

@ycombinator @kinroai @corentin_hgt @pierrealexai @parthfyi Amazing team, congrats!

[Grey Baker @greybaker]

So excited to be joining YC as a General Partner!

[Ismaeel Bashir @ismaeel_bashir_]

@nicolomagnante @yafet_melake 👀

[Y Combinator @ycombinator]

We're excited to announce Christopher Golda (@golda) and Grey Baker (@greybaker) as YC's newest General Partners!

Chris co-founded BackType (S08), sold it to Twitter, then built their entire ad business from zero to $1B+ in revenue.

Grey co-founded Dependabot, a developer tool a million engineers rely on every single day— and sold it to GitHub. Then he co-founded Pincites (S23), which was acquired by Filevine.

As Visiting Partners, both have already spent multiple batches working closely with founders. Now they're doing it full-time. https://t.co/bP6hkkvNbO

[David Lieb @dflieb]

Major congrats to Grey and Chris — both are singular talents we’re pumped to add to the YC partnership.

[Ismaeel Bashir @ismaeel_bashir_]

Early on, Jensen Huang was struggling to find a way to differentiate @nvidia

Then it hits him: he sees the OpenGL manual in a Fry's Electronics, and buys three copies. Gives one to each engineer.

what a goat. parallelising before GPUs existed :)

[will o’brien @Willob]

“So you bootstrapped the whole company?”

“Yeah, I turned it down. Turned down the funding.”

“Oh.”

“Yeah, I turned it down just to do this, just to grind it out, build it myself.”

“So you had an opportunity to take venture money?”

“Yeah, a term sheet and everything.”

“From who?”

“Yeah, a big fund out in Menlo Park. It was a top-tier firm, though. And they offered me like 15, some shit like 10 billion or something like that, 5 billion, something like that.”

“Wait, wait, wait. A term sheet.”

“Yeah.”

“Okay, but not 5 billion dollars. They offered you 5 billion dollars for your seed round?”

“Yeah.”

“For a seed round?”

“Yeah.”

“What are we doing here?”

“Like, what the y’all, like, you know what I’m saying?”

“That’s more than the GDP of a small country.”

“But I was so younger, like, I didn’t know what a cap table was.”

“Are you sure they offered you 5 billion dollars?”

“I turned it down.”

“You didn’t even have a product.”

“It was a SAFE, like, I had to give up equity for this decade.”

“But you would get $5 billion?”

“Yeah.”

“You’d be one of the most valuable companies on earth.”

“It was somewhere, it was in the billions, though.”

[Ismaeel Bashir @ismaeel_bashir_]

@dasha_shunina @ycombinator @Forbes thanks for the feature :)

[Dasha @dasha_shunina]

Meet the most promising companies from the latest @ycombinator batch 👀

One big takeaway:

AI agents need more than intelligence.
They need memory, identity, observability, compliance, insurance and power.

My latest @Forbes deep dive 👇

https://t.co/1IPkx2hqSi

[Ismaeel Bashir @ismaeel_bashir_]

Bro what? 😭 we don’t resell compute. We do increase your effective GPU capacity tho (https://t.co/hufs1EaaPJ)

These AI news articles need some validation (seems to be scraped from today’s hacker news post on us and was very poorly copied and rewritten)

Thanks for the promo tho guys 🙏😂

https://t.co/25NiNfLkig

[Ismaeel Bashir @ismaeel_bashir_]

Shoutout Expanse 😭

Gonna tell my grandkids @paulg was talking about us 🥀

[Paul Graham @paulg]

@ahistoryinart I'm puzzled by what's over the door. That would be a big expanse of glass for 1924.

[David Lieb @dflieb]

Thought experiment: if every company suddenly had infinite free compute, what new products would emerge?

My take: with very few exceptions, not much would change. The bottleneck is figuring out what people want, and it’s not so easy to apply compute to solve that.

[Ismaeel Bashir @ismaeel_bashir_]

8/

This 2× over-provisioning costs a lot, datacenters are leaving millions on the table:

1K GPUs -> roughly $8.8M/yr wasted

10K GPUs -> roughly $88M/yr

100K GPUs (frontier scale) -> roughly $880M/yr, an entire national supercomputer of capacity, idle

If you run serious compute, we would love to chat :)

https://t.co/9GXXimbOyZ
[email protected]

[Ismaeel Bashir @ismaeel_bashir_]

1/

Auto-research is taking off. Inside xAI, Anthropic, OpenAI, DeepMind, Meta SI: frontier models are running their own training experiments.

At Expanse, we benchmarked 8 frontier models to see how efficient they are at resource efficiency.

TLDR: They all ended up over-provisioning compute by 2x, even for coding specific models.

[Ismaeel Bashir @ismaeel_bashir_]

7/

When enabling Expanse, we managed to close the gap.

We achieved roughly a 10% median runtime error. And roughly 5% median memory error. About 8x more accurate than the best frontier LLM working alone.

For each deep dive above, Expanse predicted within 1-3% of truth.