Saurabh Sharma

Chief Investment Officer - AI Infra / Crypto @jump_ @jumptrading
Compiler Engineer, Quant Trader, Theatre Buff, Cornell CS grad

[Saurabh Sharma @spartaneus]

What’s up the fintechs and their Cafe’s? Saw another one from Bilt in meatpacking yesterday.. Thought it’s the new fad and then realized @CapitalOne is the OG in this game !

[Corgi @UseCorgi]

Dogpatch Cafe 🔜

[Nala @NalaZ0x]

Another one, bounty#2 unlocked $$,$$$.
​
I again used the @Kritt_AI + @Certora AutoProver workflow to isolate a critical vector that could have drained the protocol's entire TVL. Disclosed responsibly & the team was great about validating and fixing it quickly.

[Saurabh Sharma @spartaneus]

lol 😂

[MrRatable @MRatable]

Unbelievably embarrassing for Google that Gemini hasn’t committed any cybercrimes yet

[Saurabh Sharma @spartaneus]

No doubt - but still need to weather the credit and capital rotation cycles to sustain these investments.. as we recently learned being long the AI trade is the right thesis, but being levered to the max might backfire

just calling out that it is unclear what is the manageable time delta between these unprecedented investments and productive output

[Saurabh Sharma @spartaneus]

A week where AI models got cheaper and better, and AI stocks lost a trillion dollars.

Nothing broke in the technology. What broke was the financing: 400% leverage, 15-year leases against 4-year chips, $31B of capex against $784M of free cash flow. The bubble doesn’t seem to be in the models (at least yet) but more so in the capital structure !

[Saurabh Sharma @spartaneus]

Meta burned $31B of capex last quarter and has $784M of free cash flow left.
OpenAI responded by cutting its cheap model’s price 80%.

Never in history has an industry spent this hard to make its product cost less :-)

[Saurabh Sharma @spartaneus]

AI progress: teaching a model to think for 40 seconds before saying something wrong, but confidently and with citations

[Saurabh Sharma @spartaneus]

Kimi K3 - 2.8T params, ~1.5TB to download but it’s 896 experts with 16 active — only 104B params fire per token.

The largest open model ever is cheap to run and brutally expensive to hold.
Self-hosting the open frontier is now an HBM problem, not a FLOPs problem, which is why day-0 hosting went to @FireworksAI_HQ , @togethercompute and @modal .

Open weights stopped meaning open access somewhere north of a terabyte !

[Kimi.ai @Kimi_Moonshot]

Releasing the model weights and technical report of Kimi K3.

Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window.

New model architecture: 2.5x the intelligence per unit of compute, not just more params.

Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.

Model weights: https://t.co/7m7eEg6Y0B
Tech report: https://t.co/yeu6cjpMCT
Tech blog: https://t.co/YTfiMSNM1f

[Saurabh Sharma @spartaneus]

As per @FinancialTimes, NVIDIA might be the unnamed tenant on Hut 8's 1GW Texas campus. Separately, AMD just signed $14B of 15-year leases with Core Scientific.

Both are now renting data center capacity for their own chips. The scarce input in AI infra isn't GPUs anymore. It's an investment-grade balance sheet.

The trade underneath: 15-year lease duration against 3-4 year chip depreciation. Vendors are absorbing the term risk their customers can't underwrite.

https://t.co/YKsfpo9KYt

[Daksh Gupta @dakshgup]

https://t.co/VPyzZijf7t

[Daksh Gupta @dakshgup]

introducing greptile security agent

it combines static analysis, sca, and ai agents to efficiently catch security flaws before merge.

why did we build this?

- coding agents have caused the frequency of security incidents to explode

- neither agents nor deterministic methods alone are sufficient for detecting vulnerabilities

- security review is best done in conjunction with regular code review because many vulns are really just logical issues with security consequences

security review is turned on by default for free for all greptile users.

[Stani @StaniKulechov]

Aave's fintech vaults are now formally verified by Certora.

[Saurabh Sharma @spartaneus]

@armaniferrante is a rockstar builder and founder. Over the years, in all my conversations with him as an investor or partner, he is eternally optimistic and has always shown amazing conviction on Backpack’s trajectory, no matter the state of the market 🫡

[Imran @lmrankhan]

Very impressed by Armani's trajectory

Started Anchor which became the standard way to build Solana programs. Then launched Backpack directly into the FTX collapse, and then spent three years chewing glass with little to no narrative behind him

In hindsight the obvious bet was always Armani. Very few founders can go through what he and Backpack went through and still maintain the psychological endurance to keep building

I’m not an investor but the ability to absorb that many setbacks and stay motivated is one of the cleanest signals of a great founder

[Armani Ferrante @armaniferrante]

It's going to be really hard to convince startups to build stock products on Robinhood chain.

Why would you?

You're competing with Robinhood.

Robinhood chain will do great. Robinhood is awesome, but it's a different category of thing. @solana's moat is and always will be it's vibrant startup ecosystem.

The only choice serious founders have is a credibly neutral platform where your business isn't going to be directly killed by the house.

[Jonathan Claudius @claudijd]

If you’re one of these operators and frustrated with ops overhead, please let me know, we’re happy to take that off your plate.

We have a fully staffed 24/7 global SRE team led by the amazing @SEJeff.

[Lucas @variational_lvs]

We're rapidly approaching $500M in open interest on TradFi markets (making up ~1/3rd of Omni's total OI)!

Looking forward to more perp listings, the launch of swaps, and our trading API pushing this number higher.

[Mooly Sagiv @SagivMooly]

Formal verification just became inevitable.

[Jonathan Claudius @claudijd]

Our Incident Response practice is growing. We've got one opening to join a world-class team doing serious security response for some of the biggest ecosystems in web3.

My DMs are open if you want to talk.

[Sam Hogan 🇺🇸 @samhogan]

I have over 200 DMs from folks who want access to AutoTune. working to get everyone onboarded this week.

No one wants to be beholden to the labs.

The future is owning your own models and capturing the value they produce, not paying bridge tolls to the AGI cartel.

[Sam Hogan 🇺🇸 @samhogan]

We're releasing Inference AutoTune

Distill any frontier model into a 1-30B parameter task-specific SLM with only 25 lines of code

automatically route requests to reduce cost and latency by >90%

~2 hours and <$250 to train. You own the weights

Available in private beta today