Fast S-1 Teardown: Cerebras – Matt Turck

Lookup within the sky! It’s a fowl! It’s a aircraft! It’s… an IPO. The Cerebras S-1 submitting is attention-grabbing in some ways, however actually one is that, effectively, it’s an (upcoming) IPO within the first place.  In a context the place tech IPOs have been at an all time low, with a really modest uptick in 2024 (Reddit, Rubrik, and so forth), the truth that  a VC-backed tech startup has filed is uncommon sufficient to be thrilling and newsworthy by itself. 

The opposite unmistakable a part of the submitting is that Cerebras is a “pure play AI” firm, in a context the place there’s been a dearth of such firms in public markets, exterior of Palantir and arguably a few others, like C3 AI or current entrants like Tempus AI and Astera Labs. For probably the most half, public market traders have had very restricted choices to play the Generative AI wave: primarily NVIDIA, and oblique bets on AI by means of the hyperscalers. (This shortage of AI shares and to some extent, knowledge infra shares, is a actuality we captured in 2021 by means of our MAD Public Firm Index, that can quickly be price updating as hopefully extra IPOs occur).

Cerebras, rightfully so, isn’t shy about claiming the AI mantle, and by our rely, there are 137 mentions of “AI” within the prospectus abstract alone.

With the return of the IPOs comes the return of our “fast teardown” weblog posts – meant to be fast off-the-cuff breakdowns and summaries as we learn by means of S-1 filings in our areas of curiosity (knowledge, ML and AI, principally), versus any significantly in-depth evaluation.  For prior teardowns, see Klaviyo, Confluent, C3, Palantir, Snowflake.  This one is written in collaboration with my shut colleague Aman Kabeer. 

For anybody considering a bit little bit of historic context, Michael James, co-founder, had introduced at my month-to-month Information Pushed NYC occasion in February 2020: Designing an AI Supercomputer.  

HIGH LEVEL THOUGHTS

The chip wars. There’s a enjoyable David vs Goliath dimension to the Cerebras IPO story as they “come on the king”, NVIDIA (mixing metaphors from each the Bible and The Wire is one the issues we’re proudest of, in our S-1 teardown sequence). Whereas NVIDIA owns 90% of the market, there’s a plethora of different rivals each massive (AMD, Intel, Microsoft, and so forth) and rising (Groq, SambaNova, and so forth). Studying by means of the S-1s, Cerebras makes thrilling efficiency claims that ought to place it effectively from a product perspective: 10 occasions quicker coaching time-to-solution, and over 10 occasions quicker output era speeds than different high GPU-based options for inference (which permits real-time interactivity for AI purposes and brokers).

Timing is the whole lot? In fact AI is scorching, scorching, scorching, and Cerebras performs in a really strategic a part of the AI market (the well-documented AI infrastructure construct, occurring proper now), at a really delicate time (the much more well-documented GPU scarcity, which can or could not final).  Due to this fact, this certainly seems like there’s a really attention-grabbing window of alternative for Cerebras to exit, regardless of total market uncertainty (elections, and so forth), and a enterprise that’s much less absolutely baked than some would-be IPO candidates (Stripe, Databricks).

Return of the small IPO? Within the grand scheme of issues, whereas it’s rising very quick, the corporate within reason early in income – $136 million in income within the first six months of 2024, and $78.7M for the absolutely yr 2023.  In lots of conversations with each promote aspect and purchase aspect folks, we’ve heard many occasions that something beneath $400M in ARR goes to be thought of “subscale”, with some market individuals (like Philippe Laffont of Coatue) saying that $1B in ARR is the minimal quantity.  Now, many traders are sad with this state of issues, and argue that the market ought to permit smaller firms to go public – together with, curiously Brad Gerstner of Altimeter (investor in Cerebras) and Invoice Gurley of Benchmark (Eric Vishria of Benchmark led the primary spherical in Cerebras and has been on the board ever since), placing their cash the place their mouth is.  

Regardless, given the size of rivals Cerebras goes after, it is smart for Cerebras to entry the deep pockets of public markets as quickly as doable.  

Very a lot in step with the arguments of these advocating for smaller firms to go public, Cerebras could change into a great instance of an organization the place all the expansion and worth creation was not captured by non-public markets. 

Development or profitability? The market has flip flopped from a precedence standpoint over the previous few years, from ‘progress in any respect prices’ being rewarded within the 2020 & 2021 software program heyday to ‘effectivity effectivity effectivity’ in 2022 & 2023. Cerebras can be an ideal case research, significantly amidst a perceived return to favoring progress because the predominant issue.  The corporate is scaling extremely quickly (220% Y/Y from 2022-2023),  however nonetheless inefficiently (-136.3% Non-GAAP Working Margin in 2023), with a $66.6 million web loss for the primary six months of 2024, in contrast with a wider $77.8 million web loss within the first half of 2023. 

{Hardware} vs software program:  Cerebras is clearly going after NVIDIA, with spectacular product efficiency.  Nonetheless, an enormous a part of NVIDIA’s aggressive benefit lies not simply within the {hardware}, but additionally within the software program – CUDA, each a computing platform and programming mannequin enabling builders to program NVIDIA GPUs instantly.  CUDA has led to sturdy lock in inside the developer group and it’s thought of to a key element of NVIDIA’s total worth. Whereas Cerebras actually has a software program enterprise (see beneath), it will likely be attention-grabbing to see to which extent markets worth it as a {hardware} firm vs {hardware}/software program a la NVIDIA. 

Income focus:  in our teardown of C3 IPO, we had been noting how income for C3 was extremely concentrated, with the highest two clients accounting for 26% of income.  Effectively, Cerebras is taking this to an entire new degree, with 87% of Income within the first half of this yr is from only one buyer, G42 within the UAE (see particulars beneath).  Now, the character of the chip enterprise within the cloud period arguably lends itself to income focus (smaller variety of clients with deep pockets that can place very giant orders, vs each firm shopping for its personal GPU), however it will likely be attention-grabbing to see how markets react. 

ADDITIONAL NOTES

Firm / Background

• Cerebras was based in 2016 by Andrew Feldman (CEO)  and co-founders Jean-Philippe Fricker, Michael James, Gary Lauterbach and Sean Lie.  It began delivery merchandise in 2019.   
• Cerebras gives chips for high-performance computing, particularly to be used in coaching & inferencing LLMs and different fashions
  • Particular differentiation from NVIDIA and different chip suppliers comes from Cerebras’ manufacturing of ‘wafer-scale chips’ – i.e. as a substitute of breaking down a big silicon wafer into a number of particular person chips (like NVIDIA’s GPUs), Cerebras retains the total wafer intact
  • Because of this, every particular person Cerebras chip is far bigger than an NVIDIA chip (in actual fact, the most important chip ever offered), with benefits (on a per chip foundation) in variety of cores, chip reminiscence and reminiscence bandwidth amongst different issues
  • The Cerebras AI Supercomputer is an enormous cluster of those (already giant) chips, comprised of two,048 CS-3 techniques for big scale mannequin coaching & inference
• The corporate additionally gives a software program layer (CSoft) intently built-in with their chips
  • Cerebras advertises the world’s quickest inference service, as benchmarked by operating Llama3.1-70B vs. GPU-based hyperscale clouds and rising rivals
  • Cerebras’ AI Mannequin Studio as an end-to-end platform for coaching, finetuning & inference
• Relationship with G42, the Abu Dhabi primarily based Emirati AI holding firm:
  • G42 initially acquired a ~1% stake in Cerebras again in 2021
  • The S-1 reveals an settlement for G42 to take a position an incremental $335M (Collection F-2) by April 15, 2025
  • G42 already contributes nearly all of Cerebras’ Income, representing ~83% of complete FY23 Income and ~87% of 1H FY24 Income
    • That is as a part of an overarching dedication from G42 to buy $1.43B of {hardware} & companies from Cerebras, to be pre-paid earlier than February 2025
  • G42 has an possibility (the ‘G42 Possibility’ listed within the S-1) to buy extra shares at the side of giant purchases of {hardware} & companies from Cerebras:
    • If G42 purchases >$500M and <$5B in a single buy order, they’ve the suitable to buy extra most popular shares
    • Complete quantity of shares G42 should buy is calculated as: 10% of the acquisition order (>$500M) divided by both 1) if firm remains to be non-public, 17.5% low cost to the final non-public value OR 2) if firm is public, 17.5% low cost to the common public closing value over the prior 30-day interval

Financials

• Income
  • Annual
    • FY2023: $78.7M (+220% Y/Y)
    • FY2022: $24.6M
  • L6M (ended June 30)
    • 1H FY2024: $136.4M
    • 1H FY2023: $8.7M
  • Extremely unusually ~87% of Income within the first half of this yr is from only one buyer – G42 (was ~83% of FY23 Income)
• Gross Margin
  • Annual
    • FY2023: 33.5%
    • FY2022: 11.7%
  • L6M (ended June 30) – Gross Margin compression in 1H FY2024 is attributed to a reduction offered to G24 given quantity of buy
    • 1H FY2024: 41.1%
    • 1H FY2023: 50.5%
• R&D as by far the very best contributor to Opex
  • Annual
    • FY2023: 178% of Income
    • FY2022: 631% of Income
  • L6M (ended June 30)
    • FY2024: 57% of Income
    • FY2023: 881% of Income
• S&M as a a lot decrease fraction of spend, probably given focus / concentrate on one core buyer
  • Annual
    • FY2023: 12% of Income
    • FY2022: 38% of Income
  • L6M (ended June 30)
    • FY2024: 5% of Income
    • FY2023: 48% of Income
• G&A
  • Annual
    • FY2023: 13% of Income
    • FY2022: 69% of Income
  • L6M (ended June 30)
    • FY2024: 9% of Income
    • FY2023: 57% of Income
• Non-GAAP Working Margin
  • Annual
    • FY2023: -136.3%
    • FY2022: -632.8%
  • L6M (ended June 30)
    • 1H FY2024: -6.9%
    • 1H FY2023: -828.2%
• Steadiness Sheet
  • Maybe probably the most notable merchandise right here is the very giant $300M ‘Buyer Deposits’ legal responsibility, drawn from G42’s April 2024 settlement to buy a minimal of $300M price of AI supercomputer services and products
    • G24 has additionally agreed to buy an combination of $1.43B price of computing techniques, set up & help companies – not acknowledged on the steadiness sheet right now (no buyer deposit) however one to look out for on the B/S going ahead

Board & Capitalization

• Cerebras has raised $715M in enterprise capital (based on Crunchbase), most lately at a $4.1B valuation. 
• Quite a lot of well-known VC traders on the cap desk, with Alpha Wave, Altimeter, Benchmark, Coatue, Eclipse Ventures and Basis Capital all holding >5% of excellent shares
• Sam Altman is reportedly an investor 
• Board Members:
  • Worker Administrators:
    • Andrew Feldman – CEO & President
    • Robert Komin – CFO & Treasurer
    • Dhiraj Mallick – COO
  • Non-Worker Administrators:
    • Paul Auvil – prev. CFO of Proofpoint
    • Glenda Dorchak – prev. EVP & GM of International Enterprise for Spansion, fmr senior Intel exec
    • Thomas Lantzsch – prev. SVP & GM, IoT at Intel
    • Lior Susan – Eclipse Ventures
    • Steve Vassallo – Basis Capital
    • Eric Vishria – Benchmark