Patrick C Toulme

I was at an AI dinner last night in NYC.
Some interesting topics discussed

1. Performance engineers now spend most of their time managing agents and RSI loops. Most perf opts at major inference providers (together, fireworks) are found by agents.
2. RLing too much on code and math has negative effects on your model especially it’s language.
3. People were impressed by @bot and the overall acquisition of @cursor_ai by @SpaceXAI. xAI is a legitimate frontier lab right now.
4. Most people have agents in loop writing, reviewing, and submitting PRs 24/7
5. Most agreed code review is not super productive anymore
6. Consensus was Fable 5.5 will be stunning but OpenAI will beat it in December with Bel

There's been a lot of debate about AI vs human code review recently. Here's some actual data and observations from inside Meta:

- Engineers now push ~3x the code they did a year ago. Human review capacity obviously didn't triple.
- In an analysis involving 535k+ reviewed diffs, the production-incident rate for AI-reviewed diffs was 1/50th that of human reviewed.
- The revert rate for AI reviewed diffs was 1/3rd of the human reviewed.
- Median review wall time dropped by 35% compared to human-reviewed diffs.
- Experts then looked at every production incident caused by AI-approved diffs. A human reviewer wouldn't have caught any of those issues anyway in a manual review.
- Meta's internal coding agent lands ~3,000 diffs a week, and 59% need zero human edits.

Human review as default for most code changes is already obsolete, and might even be riskier.
Source in comment below.

I do not think by 2030 many large tech companies will have any code review.

This is what happens when code review is no longer a thing. You can move 1000x faster and make a much better product. assuming you have good tests/ci/verification you can commit hundreds of PRs today.

Companies with code review can never move as fast as T3 does. This is 100% the future of code.

This is what happens when you let GPT Astra/Bel + codex
1. Write all your kernels in Jalapeño ISA
2. Write your serving engine probably in Rust
3. Write your compiler + runtime
4. RL envs to hillclimb all your kernels
5. You codesign your chip architecture with Astra/Bel

The good news is that this will really lower prices for consumers and enterprises. one of the biggest benefits of RSI I think

Nice bounce but still early days for what $CBRS unlocks. Ultrafast inference is a competitive advantage in itself and Jane street deal is the proof.

Some obvious applications ultrafast is about to transform.
- computer use: every click is a model call and at today's speeds you have to watch it think after every small step. Fast enough and it'll just feel like it's working.
- voice agents/companion apps: having to wait 5-10 seconds in voice mode makes it unusable. Current workaround is to trade off intelligence for speed. Podium, another ultrafast customer is already runs it in their voice stack.
- putting LLM inference in the online path for ads and other rec systems. Think google ads personalized in real time.

There's many others. Jane street literally said: "speed (from cerebras) enables different ways of using the models" and they would not be the last customers.

Don’t sleep on @coreauto. I’m excited to see what they have been cooking.

Seems like they are 24/7 on RSI which makes sense. Hopefully they release something we can play with that isn’t nerfed. I think it’s interesting that they are a lab whose singular focus is RSI. allows them to be free of other distractions

More probable:

The high frequency trading and cyber wars of 2027 fought via low latency SRAM chips, whether Groq 3 LPX or CS-4, in disaggregated setups with Rubin.

Jane Street reportedly paying $200 million per mw - $200b per gigawatt - is likely the beginning, not the end.

CEREBRAS $CBRS: AI’S SPEED KING AND THE $1,000 BULL CASE FOR 2029–2030

My bull case for 2029–2030 come to roughly $800–1,000 per share. The financial model uses 2029, with growing customer evidence supporting the opportunity.

Start with capacity.

Cerebras’s SEC filing shows 750MW committed by OpenAI, with deployment expected during 2026–2028, plus an option for another 1.25GW by the end of 2030. That provides a contractual path to a potential 2GW. My 2029 scenario assumes an accelerated buildout and additional cloud customers bring average productive cloud capacity to 2GW during that year.

The infrastructure expansion is tangible. Cerebras announced a Finland facility scaling in phases to 165MW of contracted IT capacity, backed by seven-year service orders, with construction underway on the initial 50MW phase. This supports the expansion thesis. Achieving 2GW of productive average capacity in 2029 requires substantial execution.

On the August 12 Q2 earnings call, management said more than 600MW was live or contracted for delivery by the end of 2027, with an expansion pipeline measured in gigawatts. Manufacturing capacity was expected to increase more than 10× in 2026. Management also expects core revenue to more than triple in 2027 and continue growing at multiples in subsequent years. Those statements support the ramp thesis; $30–40B in 2029 is the bull-case projection.

Pricing?

I assume $15–20M of annual core revenue per MW, equivalent to $15–20B per GW.

There are real industry benchmarks behind that assumption. Nebius $NBIS disclosed four landmark deals at $20–25M of annual contract value per MW. CoreWeave $CRWV subsequently announced short contracts at approximately $40M/MW in annualized revenue. Those CoreWeave contracts run three to six months and reflect premium pricing.

These industry benchmarks involve different hardware and contract structures. My $15–20M of annual core revenue per MW remains an assumption Cerebras must achieve across its fleet, but it is realistic. The $25.4B backlog spans several years and includes certain pass-through amounts. Reaching $30–40B of annual core revenue in 2029 requires substantial additional bookings, timely capacity delivery and achievement of the modeled revenue yield. That being said, we are very early in the freshman year of the IPO and early in the roll out and scale up.

Market size gives this model context.

Cerebras’s IPO filings, citing Bloomberg Intelligence, estimate its addressable inference opportunity growing from approximately $66B in 2025 to $292B in 2029. My $30–40B revenue scenario equals approximately 10.3–13.7% of that 2029 forecast. At the $35B midpoint, the comparison is almost exactly 12%.

The specialized opportunity is substantial too. Public reporting of Citi $C analyst Atif Malik’s research puts fast inference at $130B in 2030. My revenue range equals approximately 23–31% of that later-year estimate, with the $35B midpoint at roughly 27%. These forecasts use different market definitions and include hardware or infrastructure alongside services. The 2030 date also matters when assessing the pace required to reach my 2029 target.

The physical footprint offers another perspective. McKinsey projects global inference demand of 71.5GW in 2029. My modeled 2GW fleet equals approximately 2.8% of that forecast. This is a nominal capacity comparison across all providers. Architecture, utilization and pricing determine the economic value of each MW and the resulting revenue share.

My thesis centers on Cerebras building a substantial, profitable position in fast inference. That specialty spans coding, agents, finance, security and scientific workloads, giving a focused architecture broad potential demand.

The customer evidence is growing too.

OpenAI identifies Jane Street, Podium, Basis and Rogo as early users of Cerebras-powered Ultrafast. Faster inference is already enabling more responsive development, voice and financial research workflows. These are customers of OpenAI’s service, powered by Cerebras capacity.

At Supernova on August 18, OpenAI’s Thibault Sottiaux said “we would love for Ultrafast to just be the default.” He acknowledged that this remained an ambition. That is a significant product aspiration from the customer already anchoring Cerebras’s capacity expansion.

SemiAnalysis (in their paywalled article we learned about during the “UltrafastGate” scandal that hit this weekend) reported on October 1 that Ultrafast service to Jane Street and a few other customers was being delivered at $200M/MW. That remains a reported outlier: I have not found independent confirmation, and the claim leaves the time period and Cerebras’s revenue share unspecified.

In his August 25, 2026 Dwarkesh Podcast interview, Dylan Patel separately discussed Jane Street capturing $300–500M/MW in economic value from AI, without documenting the calculation. My model uses the $15–20M annual core revenue per MW assumption above.

There are more disclosed commercial relationships worth including.

Cerebras signed cloud capacity agreements with Cognition and Lovable. It names Block $XYZ, Figma $FIG, AlphaSense and GSK $GSK as customers using fast inference for agentic workflows. Lovable’s agreement specifically includes dedicated Cerebras capacity for selected latency-sensitive workloads. On the Q2 call, Feldman disclosed six deals above $30M signed during the quarter.

The customer testimony helps explain why speed commands value. At Supernova, Figma’s AI research product lead Pavi Bhatter said “speed can become our strategy.” Cognition’s Silas Alberti described up to 4× faster completion of code-search-assisted tasks while preserving answer quality. These are examples of faster inference improving useful work.

General Compute announced a multi-year agreement to deploy Cerebras systems, initially targeting agentic coding. Gimlet Labs announced a Cerebras deployment for Gimlet Cloud. Callosum is integrating Cerebras silicon into its platform for heterogeneous agentic inference. These relationships expand the channels through which developers can access Cerebras performance. Their disclosed announcements provide no contract revenue figures.

Cerebras has also named IBM $IBM, Mistral, Notion and Mayo Clinic among organizations that have chosen its technology. Its SEC filing identifies G42 and Mohamed bin Zayed University of Artificial Intelligence as existing significant customers. These relationships broaden the established customer foundation.

The business is already producing revenue through both channels. Q2 core cloud and other services revenue grew 287% to $127.7M, while core hardware revenue reached $82.1M. CFO Bob Komin also disclosed several late-stage hardware opportunities representing hundreds of millions of dollars in the pipeline. Those prospective deals provide additional hardware upside as orders and revenue develop.

Cerebras earns revenue by selling systems and by selling access to its compute. Dedicated cloud contracts generally carry take-or-pay commitments, while on-demand customers pay for usage. My 2GW model assumes capacity monetized through Cerebras’s cloud services. External system sales and leases have separate economics. Systems deployed into Cerebras’s own cloud provide the infrastructure that earns service revenue, so that hardware investment and the resulting cloud revenue must be accounted for consistently.

AMD $AMD belongs in the thesis.

The announced joint architecture uses AMD Helios for high-throughput prefill and Cerebras for low-latency decode. The companies project up to 5× higher tokens per second per watt relative to a Cerebras-only configuration in their disclosed modeling. Cerebras’s Q2 update targeted production in Q4 2026.

At Supernova, AMD CTO Mark Papermaster said: “The users get better economics and low latency and that quick response.” He also described a partnership built over years, including AMD EPYC processors already used in Cerebras clusters. Arista Networks $ANET supports the networking layer, as Feldman explained at the event.

The commercial direction matters financially. Cerebras plans to buy and deploy AMD Helios systems. On the Q2 call, Feldman affirmed that Cerebras purchases the racks and receives the revenue from customers using the resulting cloud service. Discussing the joint offering, he added: “we have buyers for it already.” Buyer identities and order values were not disclosed. I found no announced AMD purchase commitment for Cerebras systems. The demonstrated opportunity is improved cloud economics and capacity to serve more customers.

Cybersecurity adds another growth opportunity.

CrowdStrike $CRWD and Cerebras have a confirmed strategic partnership. CrowdStrike will use Cerebras inference to help power Falcon AI Detection and Response. Cerebras’s Q2 update describes inline LLM security covering a large portion of enterprise traffic.

At Supernova, CrowdStrike VP of Engineering Keith Culley said fast inference enables “5x to 10x more inspection inside the same acceptable window of time.” Cerebras also highlighted Armis, now part of ServiceNow $NOW, reporting a code-security scan completed in roughly one-third the time, with more vulnerabilities found at lower cost. That comparison was presented by Cerebras; no contract value was disclosed.

My speculation is that broader deployment through CrowdStrike, larger security models and more agent activity could materially expand Cerebras compute demand. Gartner forecasts $244B in worldwide information-security spending for 2026.

CrowdStrike’s 2026 Global Threat Report found AI-enabled adversary activity increased 89% year over year, with average eCrime breakout time falling to 29 minutes and the fastest observed case taking just 27 seconds.

Inline security puts protection directly in the path of live enterprise AI traffic, making response time critical to keeping those workflows usable. Faster inference can support larger, smarter security models running continuously. That is the foundation for a potentially substantial, recurring inference business. Future deployment volumes and revenue remain undisclosed, and additional cybersecurity expansion beyond the modeled 2GW fleet remains unquantified upside.

Demand from national security customers also has a concrete foundation.

Sandia deployed four Cerebras CS-3 systems as the first part of a planned eight-node Kingfisher system, supporting the NNSA’s AI for Nuclear Deterrence strategy. Sandia, Lawrence Livermore and Los Alamos subsequently entered a third round of collaborative research with Cerebras through the Advanced Memory Technology program.

Cerebras also has a DOE memorandum of understanding supporting exploration of further AI and HPC collaboration under the Genesis Mission. Its Carahsoft partnership and inclusion in the Department of Defense’s Tradewinds Solutions Marketplace provide government procurement channels.

On the Q2 call, Feldman also described government technology partnerships involving stacked memory and integrated wafer-scale optical solutions, extending the relationship into future hardware development.

My speculation is that these existing relationships could expand into larger secure AI deployments, scientific agents, intelligence analysis and cyber defense. I leave that future expansion unquantified.

AWS and Meta add further upside.

Amazon’s AWS $AMZN has announced a hybrid inference architecture combining Trainium for prefill with Cerebras for decode, deployed inside AWS and accessed through Bedrock. Cerebras’s Q2 update targets availability in Q1 2027. That creates a distribution channel to enterprises already buying compute through AWS.

The commercial foundation is documented. Cerebras’s June 10-Q describes an initial multi-year AWS lease, options for future procurement and manufacturing capacity commitments. It explicitly says additional AWS purchases could make AWS a material percentage of revenue. My bull case includes Bedrock adoption translating into larger deployments and repeat orders as coding and agent workloads grow.

Feldman made a useful clarification on the Q2 call: “our $25 billion in RPO does not reflect any backlog of business from AWS or any other hyperscaler at this time.” That leaves hyperscaler expansion outside the backlog figure investors already know.

Meta $META also has an established connection. Its April 2025 LlamaCon announcement introduced experimental access to Cerebras-powered models through Llama API. That provides a technical precedent and an existing relationship for future expansion.

Meta’s newer Muse platform emphasizes agents, coding and reducing task latency, all areas where faster inference can matter. Broader Cerebras deployment across those workloads is a plausible upside scenario. Any Cerebras role in Muse remains speculative.

Alphabet’s Google $GOOGL deserves continued research.

Google named Cerebras as a supporter of its MLIR compiler infrastructure, and Cerebras now serves Google DeepMind’s Gemma 4 model. Those are documented technical connections. At Supernova, Cerebras CMO Julie Choi also mentioned demos from DeepMind, adding another ecosystem connection worth tracking.

On the Q2 call, management reported progress in discussions with other hyperscalers; an analyst also asked about Google and Microsoft $MSFT. I found no public Google purchase agreement or quantified Cerebras revenue commitment. I assign no Google revenue to this model.

These relationships give Cerebras additional paths to customer diversification and sustained demand into the 2030s.

The technology roadmap supports the economics too.

At Supernova, CTO Sean Lie described a CS-4 design with 50% fewer components, 60% more manufacturing automation and up to 3× faster system deployment. The roadmap targets twice the inference speed each year and solutions with up to 20× higher throughput by 2027. These are company claims and targets. Their financial significance is the potential to deliver more billable tokens per MW at lower cost as utilization and pricing evolve.

At Hot Chips 2026, Cerebras detailed CS-5, targeted for 2027, aiming for up to 5,000 output tokens per second per user on the largest frontier models. Its subsequent CS-6 roadmap adds 3D-stacked DRAM so more of each model can fit within each system, potentially reducing the infrastructure needed per workload and improving capital efficiency as those products are delivered.

My financial model:

2GW × $15–20B annual core revenue/GW = $30–40B revenue.
At 40% core operating margin = $12–16B core operating profit.
At 20× projected 2029 core operating profit = $240–320B enterprise value.
That multiple is equivalent to 8× projected 2029 core revenue.

Cerebras management has outlined a medium-to-long-term profitability target of approximately 40% core operating margin. Applying it in 2029 is part of my bull-case assumption. Core figures are non-GAAP and exclude stock-based compensation, customer-warrant amortization, data-center pass-through revenue and costs, and certain other items.

The Q2 call helps explain the margin thesis. Feldman said: “More tokens per system at lower cost means more revenue and more gross margin.” Komin reported core operating margin improving from negative 42% to negative 16% year over year, and described a path toward 60%+ core gross margin through higher pricing, improved throughput and a shift from higher-cost rented systems to owned capacity. Profitability remains a future milestone.

Capital efficiency matters just as much. Komin said Cerebras has “much lower net capital expenditure per megawatt” than most AI cloud providers. His explanation: Cerebras deploys its own hardware at bill-of-materials cost, while its largest customer reimburses a meaningful portion of data-center fit-out spending. That is management’s comparison, and it provides a concrete reason the company could retain more cash from its growth.

For valuation context, Nvidia $NVDA and Broadcom $AVGO currently trade around 28.5× and 39.8× trailing operating profit, measured using enterprise value. Their published forward P/E ratios are approximately 19.4× and 20.5×. Those figures provide context across different accounting measures and time periods.

My 20× assumption values Cerebras in 2029 using projected 2029 core operating profit, which includes equipment depreciation. At a 40% operating margin, that equals 8× revenue. The valuation requires sustained growth and cash generation after ongoing equipment and data-center investment. With an illustrative 20% tax rate and zero interest expense, 20× operating profit corresponds to approximately 25× earnings.

The midpoint is straightforward:

$35B revenue × 40% margin × 20 = $280B enterprise value.

Equivalently:

$35B revenue × 8 = $280B enterprise value.

My approximately $903 midpoint assumes 310M diluted shares and roughly zero net debt in 2029, with cash and liquid investments offsetting borrowings. The cloud buildout’s future funding mix remains uncertain. Holding enterprise value and share count constant, each $1B of net debt reduces equity value by approximately $3.23 per share.

A $250–300B equity value translates to $806–968 per share. At 360M diluted shares, it becomes $694–833. The share count and net-debt position are future assumptions that materially affect the result.

Cerebras ended June 2026 with approximately $7B of financial net cash, excluding restricted cash and operating lease liabilities. Management also describes lower net capital requirements per MW because it deploys its own hardware at bill-of-materials cost and receives reimbursement from its largest customer for a meaningful portion of data-center fit-out spending. Those advantages provide a plausible path to financing expansion through existing liquidity, customer funding and cash generated as profitability develops.

Customer usage within the 2GW cloud model gets counted once. AMD-enabled cloud usage follows that same rule. AWS leases and hardware purchases, along with systems sold to other operators, require separate revenue and margin assumptions as their economics become available.

Additional AWS, Meta, cybersecurity and government expansion remains unquantified upside here.

Capacity delivery, realized pricing, margins and dilution determine the outcome. The demand thesis centers on shorter time to decision across coding, agents, finance and security.

A prime example of a custom harness design for specific utility of the author

If you’re serious about staying at the frontier you need to be building your own dedicated harness

Every semiconductor company will need to grow by multiple times to supply a world in which every American has access to paid AI.

x.com/patricktoulme/…

Want to know how a feature works in an app, down to the binary level...?

Meet REA (Reverse Engineer Anything) an open-source toolkit that brings reverse-engineering capabilities straight to your AI agent via MCP and CLI

🔹 Decompile apps & native binaries
🔹 Trace execution & inspect evidence
🔹 Recreate features for your own stack

One of the most compelling parts about working at an ElonCo was the agility to pivot to whatever the most important problem was.

One day I said: I wasn’t hired to go after spam, but a tsunami is coming and I’m going to drop everything to focus on that.

The org is so flat and agile that your title is basically meaningless. Anyone can sound the alarm about a problem and solve it themselves—or rally the team to join them.

This is how startups work but it’s very hard to make it work in a larger company.

The opportunity cost of not burning your quota / weekly limits / tokens has IMO never been higher.

If you have access to any amount of tokens - burn them. The compute is not saved. Compute is static. It has to be powered and existing 24/7 to serve demand.

There's so much noise on the $CBRS drop this week. I think it's very misguided and this might be the best buy opportunity in a while:
- The exec sales this week don't mean much at all. They were planned months ago. And the founders are getting some liquidity AFTER A DECADE. Not a big deal.
- There were 19M shares unlocked the same day as SemiAnalysis Nvidia report. Just lots of sell pressure this week.
- As assistants such as muse/dots and voice mode become more common, faster inference will be ever more important even if gains over Nvidia are not 2.5x. Even 1.5x faster decode = 50% more thinking before it starts talking.
- There's no change to the OpenAI deal. Still 750MW, $20B+, through 2028. OpenAI also lent them $1B and holds warrants for ~10% of the company. Not a relationship you walk away from over one launch on Nvidia.

I personally have a big buy order queued for Monday morning.

Alex and Nat opened the gate to physical ai

You’re no longer confined to our form factor/wearables, any shape is now conceivable

A lot of craft went behind making the api as consumer friendly as possible

Excited to see what everyone is building

$CBRS to 1T

Since @bunjavascript was rewritten in Rust I have seen WAY less crashes in my harness.

Honestly, I wish they had rewritten it sooner. This is making me think most software should just be rewritten rather than worked around. Replaces years of tech debt in one PR.

Building harnesses and RSI loops is extremely difficult imo. Not all engineers can do it either it seems.

I do think once you know how to build RSI loops correctly you can build them for any domain

Muse who? 🏎️💨

And no surprise - it runs on JAX + XLA + TPUs. The world's best AI supercomputing stack.