It’s got five rooms (sort of) because there are always three and, well, then I added two. The map: there is Read, Ask, Founders, STAX, and a Playground.

Or you can skip the brochure and come inside:

Read is the obvious one. You read things there. It’s where From the Porch lives — research, theses, occasional rants, and whatever rabbit hole I’m currently down.

Ask the Porch is ostensibly more fun. I fed Claude an ample dose of my writing, thinking, and general Porch-ness, so you can ask it questions and get something approximating what I would say. To the delight of several friends, it also has opinions about the question: “Is Levy a douche?” This is either useful leverage or the beginning of my own obsolescence. TBD.

Founders gets more practical. There’s a founder-specific version of Ask the Porch, plus a few things I built based on posts I wrote about things I have said to founders countless times during mentor sessions. Two of those – Deck Analyzer and Stage Check – are based on posts I wrote about pitch decks and valuation. A third, Stack DNA, is a clever way to see how your tech stack matches up against other startups (and eventually could be a clever way to get your startup into the STAX database, but not now – lucky for you I don’t know how to do that yet). And a fourth (Intro Draft) would be based on this post about (people) networking – that is, if I had built it yet, which I haven’t.

There’s more to say about the Founders section. But that can be another post.

Then there’s STAX.

STAX will get its own proper introduction. I’m not going to bury the launch of one thing inside the tour of another. For now, let’s just say it’s the proprietary data layer increasingly sitting underneath Porch research that relies on the Porch Capital thesis that startup adoption of tech platforms predicts enterprise adoption of tech platforms.

More soon.

And finally there is the Playground.

The Playground serves no coherent strategic purpose whatsoever. There’s a merch shop. There’s TigerPorch and Porchlo — entirely unserious versions of two companies I tried building once before. There are interactive toys and some travel stuff. There will undoubtedly be more things that seemed funny at 1am. Not everything needs a TAM. (Also, not everything needs a playground, but sometimes you just want to share your toys.)

🥚 Easter eggs. 🐣

There are also things hidden inside the Porch. Some are stupid. Some may involve free Porch stuff. And three upstream signals are buried somewhere inside the interactive tour. Find all three and something happens. I’m not telling you where they are. That would rather defeat the point.

“It’s alive!”

This is early. The site is live. STAX is live. The founder tools are live. Other things are half-built, badly named, or sitting in my head waiting to become one of the first two. Some will work. Some will break. And some will most definitely be stupid. Historically, that has been a reasonably productive mix.

And if something makes you think: “Hmm, we should probably talk to this guy,” you probably should.

AI disclosure: I use Claude, ChatGPT, and Gemini for research, editing, and occasional editorial tickle fights. The images were made with ChatGPT – as was the interactive tour. I use Claude to build and operate STAX, which I conceived and built with Claude as my (unequal) coding partner — I don’t understand half of what it says, but I pay for it to do things for me. The ideas, analysis, and opinions are mine. To the best of my knowledge, all of the information presented is correct or eventually will be correct (with the notable exception of any sarcasm, obvious or not). Nothing here is investment advice; do your own work.

Apparently, if you chase enough unrelated ideas for long enough, eventually you’re allowed to call it a strategy. (Maybe even a business.)

Which is convenient because for the last few years, I’ve been writing From the Porch, investing in things, advising founders and funds, building weird data tools, and spending a perhaps-unhealthy amount of time thinking about GPUs, inference, capital markets and whatever bottleneck comes next.

At some point it became obvious that these were not actually separate activities. So I finally put a pretty dress on it.

Meet Porch Capital.

Porch Capital is effectively the institutional version of what I’ve been doing for a long time: trying to figure out what is happening before it becomes consensus, then finding useful ways to do something with that signal. (Oh, and trying to make a little money along the way.)

Sometimes that means writing.

Sometimes it means investing.

Sometimes it’s helping a company figure out how to position itself, raise capital, or explain what the hell it actually does.

And increasingly it means building data and tools that let me test whether the things I think I’m seeing are actually happening (rather than the murmurs from the little people in my head).

Startup adoption predicts enterprise adoption.

The core Porch Capital thesis is simple: because startups have no technical debt, make infrastructure choices faster, switch vendors on a dime, and adopt new primitives earlier, they make an ironically useful upstream sensor network for where technology markets are going.

I’ve been building some proprietary data infrastructure around this idea. The first piece is STAX — the STartup Architecture indeX. I’m going to write more about STAX separately. Because it deserves it. And also because it has to actually work. (It does, but I want it to work more.)

For purposes of this post, the important takeaway is that From the Porch research is increasingly a combination of thesis + data rather than me staring at the clouds and developing opinions whilst yelling at kids to get off my lawn. (Although there will still be some of that.)

So what did I accidentally build?

Right now, Porch is three overlapping things. (Because there are always three.)

Research.

From the Porch is where the theses start: this cycle that means AI infrastructure, the financialization of compute, agentic inference, capital markets and whatever else looks interesting, well, From the Porch. (Fifteen years ago that meant social, mobile, and cloud; dare I say Web 2.0? And twenty-five years ago it meant the Internet; dare I not say dotcom bubble — luckily for me, my equity research reports from that cycle are lost to the sands of a time when not everything left a digital breadcrumb.)

Some of From the Porch is research. Some of it is writing. Some of it is playful dreaming. Some of it is making things to play with. Increasingly I’m not sure there’s much of a distinction.

Advisory.

I work with a small number of investors, founders, banks, and operators on investment themes and analysis, company positioning, GTM, fundraising, capital strategy, the occasional “what the hell should we do here?” problem, as well as a fair bit of quasi-therapy (The jury’s out on who’s shrinking whose head.)

I’m intentionally keeping this small. I only work with people and on ideas that I like. And I don’t just take on an assignment for the money.

Unless it’s enough money, in which case my morals and ideals be damned.

Investing.

I make direct investments and occasionally put together or participate in SPVs, mostly around AI infrastructure, compute, and adjacent capital-markets infrastructure. (And the occasional healthcare / life science company that is selfishly important to me for more personal reasons.)

Aside from public positions in $NVDA, $NBIS, and $CRWV, my currently active private investments and broader venture affiliates include:

• Crusoe — vertically integrated power, data-center and AI-cloud infrastructure. I invested through an SPV in December 2024; it reflects my thesis that power and compute are AI’s core bottlenecks.

• Groq — an inference platform built around Groq’s LPU architecture. I invested through an SPV in June 2025; following its Nvidia licensing deal and shareholder distributions, I believe I’m largely out and have no financial interest in Groq 2.0; though honestly I’m not 100% sure. (FWIW I think Groq 2.0 could be very interesting; see here.)

• Throne — an AI-powered smart-toilet sensor for continuous gut and urinary-health monitoring. I’m a direct investor; it reflects my interest in AI in medicine and particularly in IBD. (A lot more to come on the IBD angle very soon.)

• Eniac — an operator-led seed venture firm backing foundational AI infrastructure, autonomous software, national security, and frontier technology.

• Factorial — Matt Hartman’s AI-focused early-stage firm, built around technical-founder sourcing partners including the founders of Hugging Face. I’m an LP in Funds I and II.

• ERA — one of NYC’s leading startup accelerators; I’ve been a venture partner and lead mentor since 2011. ERA is one of my longest-standing and closest relationships in the startup world.

• Karmel Capital — a technology investment firm focused on AI infrastructure and enterprise software through primary and direct-secondary investments. I advise on investment themes, sourcing, diligence, and related research.

The intersection is the strategy.

My career has never made much sense on a résumé. I started in banking and sell-side research during the dotcom boom. Became a hedge-fund analyst. Started two companies. Spent most of a decade around startups at AWS, with a Stripe intermission. Went back to investing. Started writing again.

I recently described myself as a decathlete rather than a specialist. That used to feel a little like a career-path defect. It turns out to be reasonably useful when the thing you’re trying to understand sits at the intersection of technology, markets, startups and capital, which AI infrastructure very much does.

So Porch isn’t an attempt to pick a lane. It’s finally putting a name around the intersection of the lanes I picked.

AI disclosure: I use Claude, ChatGPT, and Gemini for research, editing, and occasional editorial combat. The images were made with ChatGPT. I use Claude to build and operate STAX, the instrument mentioned in this post, which I conceived and built with Claude as my (unequal) coding partner — it types faster, but I control the cap table. The ideas, analysis, and opinions are mine. I accept full responsibility for everything correct; the provenance of any errors remains under active investigation. Nothing here is investment advice; do your own work.

TLDR: Stripe is a dev tools company in a payments costume. It now will own the rails for dollars and (AI) tokens. And STAX shows they run through largely the same startups.

Quick context: OpenRouter is the multi-model inference routing layer that puts hundreds of (closed and open weight) foundation models behind a single API. (It’s also the source of some awesome basic model use data in STAX. Thank you, OpenRouter; er, Stripe). This acquisition (by an ostensibly payments-related company) matters for infra because the inference middle layer – the abstraction between GPU infra and application code – just got validated (again) as a strategic asset worth owning. We’ve seen this consolidation from neoclouds like Nebius (Eigen, etc.) and chip companies like Nvidia (HuggingFace – more on that soon). And now from software (and payments?) companies.

What STAX says.

STAX can’t speak to OpenRouter’s specific adoption numbers (server side API routing leaves no public footprint, so any count here is a visibility floor), but the broader context is readable and it says that we are living in a multi-model world.

Where STAX can detect a foundation-model vendor (~19% of the corpus), OpenAI/GPT shows up at 94% of that detected category and Anthropic/Claude at 34%. Gemini and open-weight models show up in just under 10%. Note, again, that these numbers are floors given STAX (current) limitation on server-side detection. But even these scattered breadcrumbs more than hint at a multi-model reality. And OpenRouter solves the multi-model routing problem.

The (a?) Stripe angle is the (a?) non-obvious part: payments infrastructure and inference routing share the same architectural DNA: low-latency, usage-metered, reliability-critical. Moreover, Stripe has a direct billing relationship with tons of developers already paying for inference. If per-token billing becomes the norm (and I think it will), the routing layer is where the meter gets its readings.

A(nother) non-obvious angle.

Where STAX can see both payments and model layers at once, they move together. Of the 343 companies where we detect Stripe, 97 also run OpenAI and 44 run Anthropic — 1.55x and 1.97x their corpus-wide rates. In other words, a Stripe user (in STAX) is about 1.5x and 2x more likely to be running OpenAI and Anthropic than a company picked at random from the corpus. To be abundantly clear, this is correlation, not causation. Both follow from being a detectable consumer product company. What it shows is that the populations meaningfully overlap: the companies buying inference are disproportionately also the ones using Stripe to collect revenue. That’s the precondition for owning both meters.

Insider look (from 366 days at Stripe eight years ago).

Stripe isn’t a payments company. I mean, it’s a payments company. Stripe processed $1.9T of total payment volume for more than 5 million businesses (as of 2025) including Amazon, Shopify, Salesforce, and Anthropic. I can say with questionable authority that the majority of venture-backed startups that take payments use Stripe. And 88% of the Forbes AI50 already run Stripe. So yeah, they’ve made a bit of a name for themselves in payments. (All figures from Stripe’s own reporting.)

But Stripe is a developer tools company. And I mean that in the best way. To (mis)quote old Stripe colleagues: “Stripe solves problems with code.” Payments was a messy problem. Very. (Trust me; I started an e-commerce company once.) Inference routing is one too and Stripe is (now) on the case. But I can’t think of a better way to define Stripe as a developer-first company than to show this picture above of the Stripe cafeteria. It was labeled /dev/food (at least when I was there).

Watching the meter, the model, and the router – From the Porch.

Stripe bought Metronome in January. Patrick Collison (Stripe’s CEO) called it “an exceptionally powerful metering engine,” and noted it already meters OpenAI, Anthropic and Nvidia. It launched AI token billing and now it’s buying the router. That’s the meter, the model provider’s revenue, and the router in one machine packaged for developers.

What would tell me I’m wrong: if the model providers keep billing developers directly and routing stays a thin convenience layer, the meter never moves and this is just a nice acqui-product. Watch whether inference spend starts consolidating onto a single invoice or stays scattered across a dozen provider accounts. (STAX can’t see that yet either. Working on it.)

Fun fact: Stripe’s orginal name was /dev/payments. Guess or ask me why they had to change it. There may be a Porch Capital t-shirt in your future if you do 😮

Full STAX methodology at https://stax.porch.capital/#methodology

Stripe Disclosure: I worked at Stripe for a year and a day (see: vesting schedule), eight years ago. (Almost) everything above is from public sources and a corpus of >12K startups — none of it from being there. (The cafeteria sign is an exception.)

AI disclosure: I use Claude to build and operate STAX, the instrument described in this post, which I conceived and built with Claude as my (unequal) coding partner. I also use Claude, ChatGPT, and Gemini to help research, edit, and occasionally argue with the writing. The images were made with ChatGPT and the tables are from STAX and Claude. The ideas, analysis and opinions are mine. So are the jokes, unfortunately.

Financial disclosures: Long NVDA, CRWV, NBIS. Equity positions in Groq and Crusoe via SPVs. LP in Eniac Ventures and Factorial Capital. Venture partner at Entrepreneurs Roundtable Accelerator. No position in MSFT, GOOGL, Anthropic or OpenAI, at least as far as I know. Positions subject to change.

Nothing above is investment advice, including the parts that sound suspiciously like investment advice. This is research (sorta). The distinction becomes particularly important when I’m wrong.

I went looking for a fight between Claude Code and Codex. Instead, the data shows that companies are stacking rather than choosing. Which suggests switching costs are low, the fight for share is a little less consequential than it looks, and the durable value accrues a layer down to whoever sells the tokens these tools consume.

Every number above comes from breadcrumbs vendors leave — on purpose or inadvertently — like commit signatures and config files. That changed on Monday. Anthropic confirmed that Claude models launched since August 2nd now embed an invisible watermark in generated text and attach signed provenance to generated files, at the model level, worldwide. The trigger is the EU AI Act; the covered surfaces include Claude Code.

Maths, brought to you by STAX.

STAX, the Porch Capital startup tech adoption monitor, can verify a real GitHub organization for 2,416 companies. Of those, 719 carry at least one agentic coding tool, and 589 of the 719 carry Claude Code; i.e., 82% of everyone who has adopted anything.

One challenge to that 82% before you start yelling at me: Claude Code’s evidence is mostly commit trailers (498 of its 589 companies carry one) and a commit trailer is a signature the vendor chose to write into your git history. Cursor writes none, because an editor that autocompletes in place never authors a commit. Comparing them on total evidence means comparing product design and calling it market share. (Ie., don’t.)

So I decomposed it expecting the lead to collapse. But it didn’t. (“Surprise?”) On configuration files, a channel available to every tool in the category, Claude Code shows up on 352 companies against Cursor’s 116 and Copilot’s 52. Against Copilot the gap is wider on the fair comparison than on the headline. I’d have written this the other way with the same enthusiasm, which is the only reason it means anything.

Who else is in the repo?

Those 719 companies produce 1,153 detections. That’s 434 more tools than companies, so at most 434 of them run a second one, and at least 285 run exactly one. Sixty percent is a ceiling, not a count: every company running three pushes the two-tool number down.

The second tool is usually Copilot at 242 companies or Cursor at 227 — 34% and 32% of the population, close enough to call a tie. Codex is on 63, about 9%.

Co-presence alone doesn’t tell you much. Two tools on one company could just be a bake-off in progress. What rules that out is how systematic the pairings are. Companies running Claude Code are 20.9x more likely than base rates predict to also run Codex, 20.4x more likely to run Cursor, and 15.7x more likely to run Copilot. From the other direction, Codex’s strongest partners are Cursor at 28.7x and Copilot at 23.4x. Evaluations are scattered and short-lived. Twenty times chance, in the same combinations, across a population this size, is not a bake-off. It’s a stack.

Codex is the number I’d flag. It’s a small count with the strongest pairing in the category, which is a suspicious combo. STAX may only be catching it when it sits beside something already detected, in which case that 20.9x is partly measuring the instrument rather than the market. More broadly, the AI coding tool counts come from an accumulating ledger and the affinities from a live scan, and the two don’t fully agree. The ledger reads 589 Claude Code companies where the scan reads 353. No number here mixes them and I haven’t explained the gap yet.

The file that beat two of the vendors.

AGENTS.md is a plain configuration file that several assistants have agreed to honor as a shared convention. It appears on 259 companies, ahead of GitHub Copilot at 242 and Cursor at 227. (STAX tracks it separately from the tool counts so it sits outside the 719 and the 1,153 above. Same 2,416 base.)

It has no vendor behind it. It sells nothing. And more codebases have adopted it than have adopted two of the four major products in the category.

Which is the mechanism underneath the stacking. If the interface between your repo and your assistant is a portable file everybody reads, your switching cost lives in the repository rather than the vendor relationship. You can swap assistants on a Friday afternoon. (Though it may ruin your weekend.)

That’s the part I’d want to think hardest about if I were underwriting any of these companies on seats and lock-in.

What changes this month.

Everything above rests on voluntary vendor habits. Some tools sign their work in your commit history. (Nobody made them.) Some don’t.

Then Anthropic published a help doc.

Anthropic confirmed that models launched since August 2nd are watermarking generated text and signing generated files with C2PA provenance. The marking sits at the model level and applies worldwide. (Translation: Claude now signs its own work, invisibly, in everything it writes.) It’s a company decision; not something the Act requires. OpenAI has said text watermarking remains hard to deploy at scale; Anthropic has now shipped what OpenAI hasn’t. (Which raises a question about how hard it actually is.)

The trigger is Article 50 of the EU AI Act, enforceable since August 2nd, which requires providers of generative AI to mark their output in machine-readable form. Penalties run to €15M or 3% of global turnover (or, as we say here, “revenue”). Roughly 200 organizations have signed the accompanying Code of Practice, including Google, Meta, Microsoft, OpenAI and Mistral.

Drawing conclusions from the mechanism deserves care. A statistical watermark in text is a different object from a git commit trailer. I have no basis for claiming one survives into a repository or would be legible to STAX. Anthropic is admirably direct about the erasures: heavy editing, paraphrase, very short passages and format conversion can all leave no detectable mark. A detected mark tells you something was processed, but not who wrote it. And the detection tool hasn’t shipped.

Still, the direction is the point. Provenance is becoming a published, standardized, cross-vendor property of AI output, where until now it’s been a byproduct of whatever each vendor happened to do. Every external measurement of AI adoption (including STAX) sits on a floor that’s about to rise for everybody.

The caveats are load-bearing. (Or: why you should ignore me.)

The STAX AI code tool ledger only ever adds. A detection is never removed so every count above is ever-detected, not necessarily currently-using. One tool taking share from another is structurally invisible to STAX. I make no displacement claim in either direction; please don’t build a share-of-category chart from these numbers. We’ll both look wrong or stupid or both. The ledger has only run for about eight weeks so far, which limits the damage (for now), except for commit trailers, because git history never clears and they carry most of Claude Code’s count.

The denominator is narrow and it’s not neutral. STAX has enriched 11,098 companies. Only 2,416 have a public GitHub organization it could verify against their own domain. Everything here is scoped to those. Companies that maintain a public, correctly linked GitHub org skew developer-facing and open-source-adjacent. That inflates every adoption rate above. Read the levels as a ceiling and the direction as the signal, which is how I’d read them anyway.

Six tools in the category don’t get tracked at all: Tabnine, Sourcegraph Cody, Augment, Cline, Replit Agent, and Amazon Q Developer. STAX doesn’t probe for them, so they don’t appear here at all. That describes the coverage, but not their adoption.

What would change my mind?

Signing conventions converge. If Claude Code’s lead narrows against Cursor and Copilot as provenance norms harden industry-wide, a meaningful chunk of what I measured was convention rather than adoption. This is a running series and I’ll say so if it happens.

Churn the ledger has been hiding. The STAX AI coding tool ledger cannot see a company leaving a tool, so if switching turns out to be common, the stacking finding weakens a lot. I’m building the instrument most likely to falsify my own claim. Good practice or a personality disorder, take your pick.

The untracked six turn out to be big. Six named products don’t appear here at all because STAX doesn’t probe for them. If any one of them is material, the shape of the category changes.

The short version, From the Porch.

Startups are stacking coding agents, all wired to the same repo through a file (AGENTS.md) nobody owns. That’s harder to defend than a seat count suggests, and easier to monetize a layer down.

The models have started signing their own work, which means inferring AI usage from vendor breadcrumbs has a visible end date.

I’m going to miss it. (A little.)

Full methodology at https://stax.porch.capital/#methodology.

AI disclosure: I use Claude to build and operate STAX, the instrument described in this post, which I conceived and built with Claude as my (unequal) coding partner. I also use Claude, ChatGPT, and Gemini to help research, edit, and occasionally argue with the writing. The images were made with ChatGPT and the tables are from STAX and Claude. The ideas, analysis and opinions are mine. So are the jokes, unfortunately.

Financial disclosures: Long NVDA, CRWV, NBIS. Equity positions in Groq and Crusoe via SPVs. LP in Eniac Ventures and Factorial Capital. Venture partner at Entrepreneurs Roundtable Accelerator. No position in MSFT, GOOGL, Anthropic or OpenAI, at least as far as I know. Positions subject to change.

Nothing above is investment advice, including the parts that sound suspiciously like investment advice. This is research (sorta). The distinction becomes particularly important when I’m wrong.

Welcome to your stack. There’s no credit line. 🎶

Last week I wrote about the neoclouds buying their way up the stack — Nscale/Anyscale, Nebius/Eigen, CoreWeave/W&B — while the inference platforms build their way down. Together, Fireworks, Modal, Baseten. Same destination from opposite ends.

Here’s the whisper: one of those ends can borrow money and the other can’t. CoreWeave financed GPUs at 5.9%. The inference layer financed the same hardware by selling ~$4B of equity in about eight weeks. Not because one business is better — but because of what’s on the other side of the contract. That difference is worth billions, and it may decide who finishes the trip. Or who gets to build the market that fixes it.

A neocloud (like CoreWeave) building against a tech giant contract (e.g., from Meta) has a take-or-pay agreement with an investment-grade counterparty on the other side. That’s a financeable object. I wrote in April about CoreWeave’s $8.5B facility earning a first-of-its-kind A3 rating – lenders weren’t underwriting CoreWeave at all. They were underwriting Meta’s cash flows. The contracts are the credit.

Which is why the risk (or lack thereof) is whether Meta defaults. It’s not the collateral – and the chip-depreciation panic (as off-the-mark as I believe it is) misses the point. The real risk is renewal: what happens in year five when the hyperscalers’ own silicon pipelines come back online and they don’t re-up. (I’ll tell you in five years, though honestly I’m not overly concerned – ask me why offline.)

Same rack. Higher rate.

An inference platform building its own capacity doesn’t have that. Its customers range from (very) well-funded AI startups to (very) real enterprises, and plenty of that revenue is high quality. But the contracts are shorter, smaller, and rarely backed by a rated counterparty – nothing like a hyperscaler or big tech neocloud contract. A lender can’t underwrite a fifteen-year build against a one-year commitment from a Series C company. And the collateral doesn’t help: depreciating GPUs that credit committees have already been trained to fear.

So there’s a structural gap. Venture-backed AI software (e.g., inference) went from asset-light to asset-heavy in about eighteen months, and the credit apparatus that funds asset-heavy businesses hasn’t adapted. The usual suspects want covenants and counterparties that don’t exist yet in this cohort. So the AI software companies had to sell equity.

Not a nickel of it was debt. These guys are financing GPUs by selling equity at ~20x revenue — the most expensive capital on the menu, deployed against a(n ostensibly) depreciating asset. Nvidia is on three of the four cap tables, which tells you who’s still playing bank while the credit market sits it out. I called Nvidia a merchant bank back in March. Four months later it’s underwriting the inference layer’s capex through equity, because nobody else will. Again.

What the moment needs is either a genuinely aggressive credit shop or a very patient balance sheet — and there aren’t many of either.

The financialization of compute goes in steps. Steps one and two have been financing chips (hardware) and demand (contracts / projects). CoreWeave got to step two. The inference layer is stuck at step one with equity as its only option — and the only collateral it actually owns is the thing the credit market hasn’t figured out how to lend against. So could step three be financing the output? We’ve seen this evolution before in power, communications, and other markets – remember?

F*** chips. Finance tokens. (Sound familiar?)

The GPU is the wrong unit. It depreciates, its resale market is thin and volatile, and its value is entirely derivative of what runs on it. But the tokens a fleet produces are metered, billable, and — as optimization improves — produced at falling cost per unit. That’s closer to a royalty stream than a piece of equipment. Music catalogs get financed. Pharma royalties get financed. Merchant receivables get financed.

And here’s the thing: the plumbing already exists. CME and ICE announced compute futures. Ornn’s compute price index sits on Bloomberg terminals. Barkr appraises GPU clusters and Munich Re backs the valuation, so lenders can actually write against them. Ornn says it cleared the first compute swap in December 2025.

Except all of it prices GPU-hours. And Together, Fireworks, Baseten, and Modal don’t sell GPU-hours — they sell tokens. The spread between the two is their business. So for the last two years, the excuse held: you couldn’t hedge against a benchmark that didn’t measure your product, and you certainly couldn’t borrow against one.

They know the tokens are the asset. Together hired a debt capital markets banker. You don’t put a DCM guy next to your inference engineers because you’re planning another equity round.

Here’s the rub – From the Porch.

The index does exist now. Ornn launched token price indices on June 16 — built from executed, paid transactions, priced in dollars per million tokens. Silicon Data has one too, split frontier from open-weight. There is a printed, daily, third-party price for the exact thing these companies sell.

Baseten raised $1.5B of equity in June. Together raised $800M on July 1. Fireworks raised $1.505B on July 16. All of it after the benchmark went live. Not one dollar of it borrowed against it.

So the gap isn’t the number anymore. I haven’t seen anybody turn it into a financial product — no receivable, no forward, no facility. Which is the opportunity:

The first credit shop willing to underwrite committed token revenue against a published benchmark gets to invent an entirely new asset class — and price a market whose alternative is selling equity at ~20x revenue.

Meanwhile, Shanghai is reportedly designing futures denominated in tokens while CME’s and ICE’s planned contracts are denominated in GPU-hours. Two incompatible units, no equivalence framework, and the country burning 140T tokens a day is building on the one that measures output. Make of that what you will.

Everyone’s building the same company from opposite ends. Only one end has a lender.

I published the Federal Compute Bank intervention as a joke set in 2036. Two months later the CME announced compute futures. I’ve stopped making predictions about the date.

Disclosures: I am an LP in Factorial Capital, whose portfolio includes SF Compute and Modal. Long NVDA, CRWV, NBIS. Equity positions in Groq (mostly “acquired” by Nvidia, 2025) and Crusoe via SPVs. Positions subject to change. For readability, certain transaction descriptions are simplified and may prioritize economic substance over legal form. I am not a registered investment advisor and nothing here is investment advice — I’m just a guy yelling about stuff From the Porch. Do your own work.

Welcome to your stack. There’s no turning back. 🎶

The AI stack is collapsing into one company, and both ends are racing to build it. The neoclouds are buying their way there. The inference platforms are building theirs. Only one of them can borrow the money — but that’s Part 2.

Neocloud newcomer Nscale bought AI orchestration software specialist Anyscale last week in what was at least the fifth “AI infra buys AI software” deal in the last year(ish). The nerds with pagers and GPU clusters are buying the nerds with hoodies and 2am commits that write the software that decides what the clusters do.

Anyscale is the company behind Ray, an open source AI workload orchestration framework its founders created at Berkeley. According to Bloomberg, Nscale is dropping ~$1.65B for Anyscale’s commercial platform, engineering team, and customers like Coinbase, Runway, and Bedrock Robotics. Ray went to the PyTorch Foundation in 2025 and stays open source. And Anyscale gets to keep its now superfluous “A” and “y” as the Anyscale-branded AI orchestration arm of Nscale.

This isn’t an isolated move. Nebius paid $643M for Eigen AI and IREN bought Mirantis in May. CoreWeave bought Weights & Biases in 2025. And Qualcomm bought Modular seemingly five minutes ago. Five deals, one direction: whoever owns the metal wants the layer that decides what it does.

Fwiw, I wrote about this trend here and (literally) said: “Another AI infra deal. Except it isn’t. And skipping it is a mistake.” But I say a lot of stuff so I’m bound to get something sorta right on occasion.

The neoclouds bought the ops layer.

The STAX startup adoption monitor from Porch Capital tracks ~12K VC-backed company stacks. Data from the MLOps category is thin to say the least (58 companies, ~0.5% of the base). But look at who’s inside that half percent:

Ray and W&B are 60 of the 68 signals in the category. (Note: some companies run more than one of these.) As of July 30, both are spoken for — CoreWeave has one; Nscale, the other. The entire commercially meaningful MLOps layer in the STAX base changed owners inside eighteen months. The deals were the news. The category is the story.

(Side note: the STAX MLOps data is so thin because almost nobody at the application layer runs MLOps tooling — at least not visibly. Which is itself a finding: startups consume models, they don’t operate them.)

What they’re actually buying.

Everyone calls raw GPU rental a commodity with a price sheet and no moat. That’s lazy — power contracts, interconnect topology, and delivery dates are real moats, and anyone who’s tried to site a datacenter knows it. But it’s directionally true enough that every neocloud is acting on it.

The software layer is where you stop competing on dollars-per-hour and start competing on dollars-per-finished-job — and where a customer, once integrated, doesn’t leave over a spreadsheet. Own both layers and you can co-design them, which means you capture the efficiency gain instead of handing it to the customer as a lower bill.

Which is why the scheduler is worth $1.65B to somebody who sells GPU-hours. It’s the layer that decides how many hours the job takes.

Now flip it.

While the AI infrastructure companies buy software, the (AI) software companies are building infrastructure. And one, Lightning AI, merged with GPU provider Voltage Park in a $2.5B deal — with Lightning as the surviving company.

Together, Fireworks, Modal, Baseten — the inference and serving layer — sit somewhere on a spectrum from fully leased to increasingly owned. Some are staying asset-light on principle. Others are building. The strategic destination is the same as Nscale’s (and CoreWeave’s and Nebius’s and….): own the metal and the software layer that orchestrates it.

Same destination, opposite directions – From the Porch.

Everybody wants to rule the whole stack. The neoclouds are buying their way there. The inference platforms are building theirs. And one of those roads is a lot more expensive than the other.

Disclosures: STAX is built and operated by Porch Capital. STAX data referenced here covers ~11,820 VC-backed company stacks as of the August 3, 2026 scan — directional signal, not a census. Detection is based on publicly observable stack signals and understates private or self-hosted deployments. I am an LP in Factorial Capital, whose portfolio includes SF Compute and Modal. Long NVDA, CRWV, NBIS. Equity positions in Groq (mostly “acquired” by Nvidia, 2025) and Crusoe via SPVs. Positions subject to change. For readability, certain transaction descriptions are simplified and may prioritize economic substance over legal form. I am not a registered investment advisor and nothing here is investment advice — I’m just a guy yelling about stuff From the Porch. Do your own work.

For “forever” the constraints on AI infrastructure were a tidy technical list: chips, power, memory. But the AI build-out’s latest bottleneck is the communities where these power-hungry and (allegedly) cooling-water-thirsty token factories are popping up.

Solution: data center builders are hiring community-affairs managers, government-relations reps, and public-policy directors as fast as they can staff them. Whole departments in some cases. Two years ago most of them didn’t have this function at all. Now, you can’t swing a dead cat at a neocloud job board without hitting a community-affairs req.

Teachers and tacos get $50B from Meta.

This week, Meta announced it increased its rural Louisiana data center investment to >$50B. Meta’s press release didn’t lead with gigawatts or model training. Instead, the headline (from Meta) was “Teachers and Local Businesses Win as Meta Expands Louisiana Data Center.” And that $50B bought them quotes like:

“Our roots are in this community — and Meta gave us the courage to grow here.”
-Tim and Lindsey Allen of Holy Tacos in Rayville, Louisiana.

And:

“It’s life-altering for our teachers and their families, and it’s transforming our schools. Meta’s investment has made Richland Parish a destination for education as well as industry.”
-School District Superintendent in Richland Parish, Louisiana.

None of this is charity. (Duh.) Meta suddenly transformed into a community organizer with a $50B checkbook and a comms team. But the instinct is exactly the one lighting up those job boards: win the neighbors before the neighbors organize against you. Meta’s just rich enough to buy the goodwill wholesale. Most builders have to hire for it.

Which raises the question that the press release is built to keep you from asking: what is a company Meta’s size so worried about that it opens a data-center announcement with taco testimonials?

NIMBY, meet the AI build-out.

You already knew about this. A hyperscale campus draws as much power as a small city and residents (justifiably) fret about their utility bills. Data centers allegedly drink water for cooling in places that don’t have it to spare. That said, builders are quick to note that (1) watering a single golf course is more taxing on liquid resources and (2) data centers employ a closed loop water system so they’re not truly ingesting new water.

Right or wrong, the community perspective is what it is. These data centers hum, run diesel or gas backup, fill two-lane roads with construction traffic, and often get generous tax abatements in exchange for a few dozen permanent jobs. Stack that on top of the sheer pace and scale of the current build-out, and you get organized resistance.

In the first quarter of 2026, Data Center Watch counted ~$130B in projects blocked or delayed by local opposition across 49 states — nearly matching all of 2025 in three months. And a Gallup poll said 70% of Americans oppose having an AI data center built near them — higher than opposition to a nearby nuclear plant. (Sidenote: I’m negotiating the movie rights to Silkwood 2 with Anne Hathaway as the new Meryl Streep.)

AI build-out, meet political capital.

Never missing an opportunity to capitalize on real or fabricated public outrage, statehouses are drafting legislation. And Bernie Sanders penned the AI Data Center Moratorium Act to halt the construction and expansion of AI data centers. Natch.

That’s the wall. Meta’s teachers-and-tacos rollout is one way over it. Staffing up community-outreach and public-policy roles is another — and that one leaves a paper trail. Which is convenient, because job reqs happen to be a spin-resistant way to track this whole phenomenon: they draw a cleaner picture of where the industry is getting tricky than any press release will.

The job req tea leaves – from STAX.

Using STAX (Porch Capital’s startup adoption signal tracker), Claude and I (and STAX) pulled the live boards for ~50 companies across the AI-infrastructure stack, filtered the titles to the community-affairs / government-affairs / public-policy family, dropped the false positives (a “community manager” at an inference startup runs a Discord, not a zoning board).

The results show just how much ground a company breaks on its own. Unsurprisingly, neoclouds are where we’re seeing the most (new) action. Here’s the live count of community-affairs, government-affairs, and public-policy postings I found, grouped by tier (sort of).

You can call it directional (generously), thin, biased, incomplete — pick your adjective. And yet the pattern survives: build gigawatt campuses and you staff this function; lease or distribute across other people’s buildings and you don’t. Two things jump out. Nscale — emerging enough that name-dropping it at a dinner party is an AI nerd flex — has six of these roles open across five countries, wiring the function into its foundation as it pours concrete. And the work’s gone senior and global: Equinix is hiring a VP of Public Policy; Vantage, a Senior Director for APAC. This is no longer someone to host the open house.

It’s also spreading geographically: Virginia’s Data Center Alley owned this story for years, but fresh postings pin two separate companies — Nebius and QTS — to Pennsylvania, one of them to a real, contested project in Salem Township. When two independent operators plant flags in the same new state, a front is opening.

One honest caveat, because the tempting version of this piece is a chart of reqs exploding in 2026, and that chart would be half a lie (and half a truth). What I pulled is a snapshot of live postings; expired ones vanish, so recent roles are just easier to find. The real signal isn’t the raw count — it’s the founding dates, which don’t depend on a posting still being open. This function barely existed at the neocloud tier before 2025. CoreWeave stood up its function in May 2025, hiring a VP of Government Affairs. Nscale seems to be doing it now. That’s the phase change, and it’s the part I’d stake the thesis on.

What it costs, and what it’s worth. 💰

The investor translation is short: this function is social license as capex — a cost line and a moat wearing the same jacket. Builders who arrive of a community, not merely in it, clear zoning faster, and that runs straight to build timelines and cost of capital.

Which reprices a few things. The neoclouds racing to staff this are buying down an execution risk the market hasn’t marked yet — they’re not just adding overhead. The asset-light shops that look leaner haven’t escaped that risk (yet). And Meta spending part of $50B to make a parish’s teachers the highest-paid in the state, with the stock down on the year? Cringe if you want, but that’s the cover charge for deploying at hyperscale.

The alternative is Memphis, where xAI built in the dark, bypassed the municipal grid to hook up dozens of unpermitted gas turbines, and triggered a wave of environmental lawsuits from local communities. But even the “move fast and break things” crowd has to pay the tax eventually: pull up xAI’s job board today and you’ll find a fresh listing for a Permitting Specialist in Memphis. The required skillset? “Experience with... noise assessments and community engagement.” Elon may be able to influence (buy?) an election (or not), but he still has to hire to cozy up to the neighbors.

Community organization, From the Porch.

Here’s the fix for the snapshot problem: STAX will start to archive this signal weekly — a forward-only record of the build-out’s soft underbelly, assembled one org chart at a time. Check back in a quarter and maybe I can show you the slope instead of asserting it.

And spare a thought for the punchline that wasn’t. In 2008, a VP nominee got a convention hall cackling at the idea that being a community organizer was anything like real work. Lipstick, pit bulls, hockey moms… the whole bit. Eighteen years later, the fastest-growing and most capital-intensive industry on Earth is paying six figures for exactly that skill: going into a community, listening, and building enough trust to get something hard actually done. The once-mocked community organizer turned out to be the hire you can’t build without. (And also turned out to be one of the most consequential presidents in American history; and, disclosure, my favorite.)

The industry learned it the expensive way — the way industries always do. If you want to know where it’s about to get hard next, don’t wait for the lawsuit or the moratorium. Read the reqs. They name the walls before the walls make the news.

Disclosures: I’m long CoreWeave and Nebius — yes, two of the names I just talked up — and may still hold Groq via an SPV, which even I can’t fully diagram. Nothing here is investment or political advice.

Last week, reports surfaced that Meta is planning to rent AI compute through a newly minted Meta Compute unit. Of course CoreWeave, Nebius et al got pummeled as investors shit their pants about a new mega competitor and existing and future Meta neocloud contracts.

Possible? Of course. Market overreaction? Probably.

But before jumping to either conclusion, let’s travel back to 2013 when Facebook (as it was called then) acquired Parse — and one (unnamed) AWS startup BD manager nearly shit his pants. (Spoiler alert: that was me.)

Facebook ❤️ Parse.

In 2013, Facebook spent ~$85M to buy Parse, which was one of the hottest startup developer platforms — a Backend-as-Service that gave mobile devs database, auth, APIs, push notifications, storage, and cloud code and let them eschew building or managing any infrastructure. It powered ~100K apps then and >500K a year later. It was Facebook’s Trojan horse to entice developers to integrate Facebook Login and Analytics to feed data back into its ad-targeting machine.

By the time I joined AWS in 2014, Facebook’s Parse lurked around developer circles waiting to feast on our startup lunch. Most scary (and valuable) was Parse’s relationship with developers. Our team’s MO was to own that startup relationship early because we knew that “today’s” startups are tomorrow’s enterprises.

Long before STAX (the startup signal tracker from Porch Capital), I was obsessed with startup adoption predicting enterprise adoption. If Facebook became the default platform for startups, who knew where that could eventually lead?

As it turned out… nowhere.

Facebook 💔 Parse.

In 2016, Facebook announced it was shutting down Parse — presumably to focus on its core advertising business and avoid a capital intensive battle for infra marketshare. (Anecdotally, the Parse shutdown presented an opportunity for us at AWS to pick up all of those soon-to-be infra homeless startups. Seems like “most” went to Google’s Firebase.)

Parse reminds us that just because a company can build infrastructure doesn’t mean it wants infrastructure to become its business. Which brings us back to Meta Compute. The debate isn’t whether Meta can compete with CoreWeave. Of course it can.

The question is whether Meta Compute is the beginning of a serious cloud business — or “simply” a way to generate incremental returns on an AI infrastructure investment that already measures in the hundreds of billions of dollars.

Infrastructure isn’t Meta’s product. It’s an asset.

AWS and CoreWeave wake up every morning trying to acquire compute customers. Meta wakes up every morning trying to build the best AI products in the world (and trying to sell a few ads based on our most personal and scintillating information). That’s an important distinction.

The SemiAnalysis team published an excellent deep dive on what Meta Compute could become. I recommend reading it and I largely agree with their conclusion that Meta’s expanding AI infrastructure ambitions include both building and buying infra capacity.

But if you’re spending >$100B a year building AI infrastructure, a natural capital allocation question is whether this asset can generate returns when it’s not in use.

Meta isn’t the first to ask this question. SpaceX is monetizing excess AI infrastructure. Enterprises are discussing how to lease idle GPU capacity. SF Compute exists because companies overbuy compute and then need a marketplace to sell or sublease it. If plenty of companies are already doing these distributing networks of idle compute, maybe Meta Compute is simply the largest expression of that trend yet.

The biggest change now, however, is this:

Owning GPUs used to be a cost. Increasingly, it’s an income-producing asset.

Compute is capital.

I’ve written about the financialization of compute. Not because GPUs suddenly became valuable — at ~$50K per Nvidia GB300 chip that’s hardly a debate — but because compute output is becoming something that can be bought, sold, reserved, financed, subleased, and monetized across a growing ecosystem.

Once you stop thinking of GPUs as (costly) hardware and start thinking of them as factories producing compute, a very different world begins to materialize.

• Infrastructure can generate income.

• Future compute production can be sold forward.

• Idle capacity can be leased.

• Compute credits can become a medium of exchange between partners.

• Long-term compute contracts can support financing.

• Marketplaces can match buyers and sellers.

AI capital allocation, From the Porch.

Maybe Meta Compute becomes a major cloud business. Maybe it doesn’t. Frankly, that’s almost beside the point. I think investors should fret less about whether Meta is “becoming CoreWeave” (it’s not) and focus more attention to the broader shift underneath.

We’re watching AI infrastructure evolve from a cost center into productive capital. And once that happens, monetizing compute is just good capital allocation.

And the companies that own the next generation of AI infrastructure may discover they’re operating factories that produce one of the world’s newest financial assets: compute.

Disclosures: I often invest behind the themes I write about. As of publication I’m long CoreWeave (CRWV), Nebius (NBIS), and Nvidia (NVDA), and hold Crusoe Energy via an SPV. I may add to, trim, or exit any of these without notice and won’t update this post when I do. Nothing here is investment advice or a solicitation — do your own work.

For me, it’s been a long time since building felt like playing. But that’s what building with AI has started to feel like. It feels a little like someone handed your imagination a 3D printer.

So I decided to test that theory. Over a few weekends, I (i.e., Claude) resurrected two startups I buried more than a decade ago. 

TigerBow was my first — b. 2007, d. 2010; and Philo was my second — b. 2010, d. 2011. (Founder career: still dead.) They still probably shouldn’t exist. Which is exactly why rebuilding them taught me something I wasn’t expecting. 

The two ghosts. 

TigerBow let you send a real, physical gift to someone using nothing but their email address. No shipping address, no awkward: “Hey what's your mailing address?” text that ruins the surprise (or alerts the authorities). You picked a gift, TigerBow handled the double-blind matching between the virtual you and the physical them, and a box showed up at their door.

It was a neat (and kinda creepy) idea. It was also a nightmare to build in 2007. We had no clean product catalog. Checkout integration with retailers was a custom slog for every partner. And we spent most of our runway trying to prove a value prop that needed scale to feel magical. Then Amazon and everyone else quietly commoditized gift-to-email, and the window closed. (A startup called Goody built this exact thing a decade later — same no-address flow, same everything — grew 5,000% in two quarters and raised ~$32M from NEA and SoftBank. So the idea was fine. The timing was bad. And my execution was worse.)

Philo was social TV (specifically, foursquare for TV) — a second-screen companion that let you check in to what you were watching and talk about it with other people watching the same thing. (Not to be confused with the — much more successful — Philo that exists today. Different company. We were first to the startup name and first to the startup graveyard.)

The product itself wasn't the hard part. What killed us was that Twitter and Facebook won the “talk about TV in real time” behavior before we could. You don't build a social network next to the social networks. We were directionally right and structurally doomed. Fwiw, I hold the dubious honor of having the highest dollar return of the early TV check-in companies. (Reason: I only raised $1M+ while my closest comps raised $24M and $6M. We all failed. So this is my version of “winning.” 🏆)

Two honest post-mortems (here and here). Both ideas were early. Both teams were under-resourced. Both companies were the kind of swing you don't regret taking.

Two AI resurrections. 

TigerPorch is TigerBow, rebuilt. Sort of. (It’s real. For real.) Same core promise — send a real gift to an email address, no shipping info required. This version takes (real) money through Stripe, pulls from a real catalog, and ships physical (real!) product through a real fulfillment pipeline. It's not a mockup. You can send someone a thing right now and a thing will arrive. (Unless it breaks. But it probably won’t.)

Porchlo is Philo's spiritual rebuild, borrowing from music instead of TV. Spotify OAuth, a recommendation quiz, a live (fake) leaderboard, custom domain, the works. It's the lighter of the two — more toy than business — but it shows the range. (PS — it’s also real. For real.)

Both went from nothing to live, working, end-to-end. The thing that took us months and a $1M+ VC-led seed round in 2010 took a handful of weekends between everything else. (Note: $1M+ was a respectable seed round back then. 👴)

You’ve already heard this story: AI made building stuff wicked fast.

It’s true. Payments, auth, email, fulfillment, DNS — things that once took teams and months now come together over a weekend. Most of the work isn’t writing code anymore. It’s deciding what should exist.

The friction just moved. Every “I built an app in an afternoon” post skips over the hours spent chasing broken deploys, mysterious API errors, and tiny details that don’t quite match the picture in your head.

AI didn’t know “no shipping address required” was the soul of TigerBow (or TigerPorch). It didn’t know the magic of Philo (or Porchlo) was giving people something fun to do together — not maximizing engagement. AI will happily help you build the wrong thing — beautifully and very quickly. (Trust me. 😳)

For founders and investors alike, a product demo and execution are table stakes. Judgment — knowing what deserves to exist — is the scarce resource.

A 3D printer for your brain, From the Porch. 

It's not just that building got cheaper or easier or faster. It's that building got possible — for things that used to live permanently on the other side of “I can't code that.” Ideas I would’ve abandoned in 2007 because they weren't worth a contractor and a quarter are now just… reachable. You think of a thing, and you can make the thing. The distance between an idea and an object collapsed.

It's almost like someone handed your imagination a 3D printer. (Note to self: get 3D printer.) Whatever you can picture clearly enough, you can have a working version of by the end of the night. 

The constraint was never imagination — it was the brutal cost of getting from imagination to a thing that runs. That cost just fell through the floor and what rushes in to fill the space is play.

P.S. TigerPorch and Porchlo are real, functioning applications. They live at the Porch Capital Playground, which has become the 3D printer for Porch Capital’s imagination.  

P.P.S. If this article makes you build something over the next couple of weeks, tell me what it is. I’ll randomly send a few builders a Porchlo or TigerPorch tee. No shipping address required.

That AWS may sell Trainium chips to companies not called AWS isn’t entirely surprising (at all). The demand for AI silicon is insatiable and AWS can make a bundle of money and expand the ecosystem of customers that want to build on Trainium and ultimately be more likely to use… AWS. (See: flywheel.)

But there’s a very Amazonian nuance to why AWS might sell chips externally. It’s this:

Amazon knows that its businesses have to use the best tools available, not tools protected from competition. That’s one reason Amazon turns internal capabilities into products. Once customers, including Amazon itself, can choose them — or reject them — the teams behind those products are forced to compete, improve, and innovate (or die). The result isn’t just a new revenue stream. It’s a mechanism for making the product better, which ultimately benefits both customers and Amazon’s own businesses.

Amazonians (past and present) will recognize the pattern. Amazon has long pushed its internal products to earn adoption (elsewhere internally) rather than assume it. The goal is for Amazon’s business units to use the best tools available, not just tools Amazon makes. Exposing a product to external customers creates the competitive pressure that helps make that happen.

Exhibit A: AWS had to earn it.

To find another example look no further than AWS itself. Amazon built infrastructure to solve its own compute resource problems. But once that infrastructure became a product, it had to compete for customers every day. That market pressure transformed an internal capability into one of the most important technology platforms in history.

Insider BS: when I was at AWS the story was that Prime Video originally used Akamai’s CDN over Amazon’s own Cloudfront CDN because of Akamai’s superior performance. True? Considering Akamai was doing video streaming before Cloudfront’s release in 2008 it either calls bullshit on the story or strengthens the thesis. Regardless, I can assure you that whichever Amazon manager carried the Prime Video usage goals had zero interest in risking Prime Video’s metrics to support a fledgling AWS product.

Fast forward to as recently as last year, Akamai was still the #1 CDN for enterprise with ~31% share, but Cloudfront went from 0% (nonexistent) in 2008 to just about tied with CloudFlare for the second spot at ~21%. Porch Capital’s STAX signals monitor tells an even more compelling story. For venture-backed startups, which I have repeatedly suggested predicts enterprise (and overall market) adoption, Cloudfront is in the clear #2 spot with 35% of startups tracked emitting signals of CDN usage. CloudFlare has jumped from nowhere to 53% (and the pole position) while Akamai barely registers at all.

(Fine print: STAX tracks startup adoption, not market share. It’s built from public signals and breadcrumbs, so it’s imperfect by design — but often directionally useful.)

Exhibit B: Another cost center becomes a product.

The same pattern played out in fulfillment and logistics. Amazon built warehouses and fulfillment systems because its retail business needed them. Those capabilities eventually became Fulfillment by Amazon (FBA), allowing third-party merchants to leverage Amazon’s infrastructure. As volume exploded, Amazon expanded deeper into logistics, building sortation centers, trucking networks, and eventually Amazon Air. What began as internal capabilities evolved into products and services used by others.

Once external customers entered the picture, those organizations weren’t just serving Amazon anymore. They had to compete. Every merchant became another source of feedback, pressure, and improvement. The result: a new revenue stream; and it was a better fulfillment and logistics platform for customers — and for Amazon itself.

Today, ~60% of everything sold on Amazon is sold by third-party merchants, and >80% of those sellers use FBA. What began as infrastructure built for Amazon’s own retail business has become one of the largest fulfillment platforms in the world.

From the Porch: The market gets to decide.

I’m guessing most people saw the Trainium buzz as Amazon is fighting back against NVIDIA’s stranglehold on supply of chips AI desperately needs; as well as a move toward vertical integration like Google has with cloud and TPUs (and Gemini). Maybe. (I.e., probably / definitely.) And those people aren’t wrong.

But there’s also this slightly different lens to keep in mind. AWS itself was once an internal capability. So was CloudFront. So was FBA. So was much of Amazon’s logistics network.

Trainium could be the next internal capability Amazon is exposing to the discipline of the market. And, with that lens, its goal is to create the conditions under which Trainium can become a better chip platform.

Whether it happens remains to be seen. But if history is any guide, Amazon’s willingness to sell Trainium may be less a statement about monetization and more a statement about confidence.

Amazon believes the product is ready to compete.

Now the market gets to decide.

Disclaimers and disclosures: I spent most of 2014-2023 at AWS. Nothing in this post is based on confidential information, internal documents, or non-public discussions. These are my observations as a former employee combined with publicly available information. I do not currently own any individual shares of $AMZN (that I’m aware of). Fwiw, overall, Amazon was an amazing place to work. It had its moments, but it shaped the third or fourth incarnation of my career into something that I can’t easily put into words. (To Amazon Legal: if I’ve accidentally disclosed anything confidential, it would be the first time in history that a former Amazon employee voluntarily wrote a disclaimer this long.)

In 1977, Bill Gates and Paul Allen now infamously envisioned (something like) “a computer on every desk and in every home, running Microsoft software.” Twenty years later, that vision was reality.

By 1997, there were roughly 100 million computers in American homes. But only about 20% of them were connected to the Internet. To a younger, even more insufferable version of myself working as an Internet analyst on Wall Street, that gap looked destined to close.

Why?

Because every new computer sold in 1997 was being sold with and for the purpose of connecting to the Internet. And as it turned out, the gap did close. Not just in the United States, but across much of the world.

Which raises a suspiciously déjà-vu-y question: Is Jensen Huang trying to do the same thing?

Nvidia recently announced a new generation of AI-focused Windows PCs and previously launched its DGX Spark workstation, a compact AI supercomputer for developers.

Most people looked at these announcements and shrugged. (Most people = me.) The (my) first reaction was:

Fair question. After all, Nvidia’s new Windows AI PCs don’t look much less capable than the $4,500 DGX Spark workstation – similar silicon, up to 128GB of unified memory, and ~1 petaflop of AI performance. Same guts, new shiny box? 🤷‍♂️ But before we answer who, we have to ask why.

So here’s why….

Based on my extensive research (i.e., none), I submit that the majority of consumers believe that AI (i.e., ChatGPT) is either free like Google or $50-100/mo unlimited like their cell phone bills. A more ironic analogy: it’s like their $20/mo AOL accounts. Why? Because in 1996 AOL switched from usage-metered pricing to $20/mo for unlimited access. That changed the way we consume most of our digital products from metered service – accompanied by the occasionally wallet-melting surprise monthly bills – to safe, predictable flat-fee for unlimited access. So for three decades we’ve been conditioned to pay flat monthly fees or use (ad-supported) services for free.

The Meter Starts Running (again)

The all-you-can eat digital buffet perception breaks down quickly once people scale the consumer AI adoption curve. (You can also see where you are on the curve on the interactive AI adoption curve thing I made — because clearly I’m at least at stage 4.)

My first confounding experience with metered AI was Manus, the AI agent that Meta acquired for $2B before China made them “unacquire” it. I wasn’t building complicated financial models or multimodal agentic AI applications. I was using it for three very consumer-y activities:

(1) Making personalized travel guides / itineraries for family trips (they’re pretty sweet – you can see a few here.)

(2) An agent to check daily refund status of Dubai family vacation flights that I booked literally days before the (most recent) war in the Middle East started. (It was my second time booking on Emirates. The first was a minute before Covid. Wanna know when I book Emirates next? 😱)

(3) An agent to collect my weekly todos from Granola notes. (To be fair, this one was a little more complex than a simple consumer app, but still.)

Creating travel guides (at least the way I make them) burns tokens pretty quickly — lots of graphic design. And flight scheduling and todo list agents that refuse to stop running will gobble their way through a once seemingly endless supply of Manus tokens.

These projects were my gateway drug to Claude. I’m not sure I made it one day on the $20/mo plan. Building porch.capital and STAX almost immediately demanded at least Claude Max 5x and I still hit usage limits and the occasional overage.

Point being, even now you don’t need to be a software developer to require more than a cheap, all-you-can-eat subscription. And it’s not a stretch to imagine everyday consumers increasingly relying on AI to help plan, organize, remember, monitor, and automate parts of their daily lives. That will inevitably require a more expensive subscription that comes with usage caps and metered service.

And if millions of households start running AI continuously rather than occasionally, the economics start to look very different. Consumer AI bills start to look very big. That’s when the idea of having some AI infrastructure closer to or in your home starts looking a lot less crazy — because the incremental cost of using AI@Home is close to zero when you’re not paying a metered rate dependent upon token consumption. (We can talk about electricity bills another time.)

OpenClaw ❤️ Mac mini.

The first people to notice hints of AI@Home were developers (not consumers). Earlier this year, OpenClaw unexpectedly turned the humble Mac mini into the hottest AI appliance on the planet. Developers and power users discovered that a $599 Mac mini could sit quietly in a closet, run AI agents 24/7 via open-source frameworks like Ollama, and avoid many of the usage limits and metered costs associated with cloud / token-based AI. It turns out Apple’s unified memory architecture makes it a powerhouse for running local LLMs. Demand surged. Supply was tight. Used Mac minis that normally depreciate started selling for hundreds of dollars above retail, with some configurations approaching $1,000 on secondary markets.

To be clear, we’re talking about developers and AI enthusiasts, not mainstream consumers. But the people filling their homes with AI appliances today look a lot like the people who were installing Wi-Fi routers, home networks, and broadband connections before everyone else.

And Apple itself, the undisputed deity of all things consumer electronic, can’t stop talking about on-device AI, local processing, personal context, and models running on your own hardware. Politely, they haven’t quite gotten there yet, but they swear this new Siri AI is it. (Again.) More than a few startups have been working to crack the AI-on-device nut including the team at Mirai Labs, which fancies itself the frontier on-device AI lab, building the models, inference runtime, and quantization stack so AI can run at full capability on the hardware billions of people already own. (I copy / pasted at least half of that so it must be right.)

It’s not just technology companies thinking this way. Earlier this year, Nvidia partnered with homebuilder Pulte to explore what “AI-ready” homes might look like. Whether that ultimately matters is beside the point. The interesting part is that the conversation has already moved beyond devices and into infrastructure. At some point, people stopped asking whether homes needed Internet access and started wiring it into the walls. The whole thing feels a little suspiciously déjà-vu-y all over again.

Nvidia suddenly discovers CPUs.

This could be why Jensen Huang suddenly discovered the CPU market. “Vera opens a brand new $200 billion TAM for Nvidia, a market we have never addressed before…” he gleefully declared about his new CPU on a recent earnings call. This from the guy that in 2009 called the debate about CPUs (from frienemy Intel) and GPUs (from Nvidia) “the battle for the soul of the PC.” I suppose everybody wins now. Or at least Jensen.

Don’t get me wrong, these monster supercomputers are going into data centers and warehouse-sized AI factories. But the same technology is now showing up in a new generation of Windows PCs that Nvidia describes as platforms for “personal agents” — designed to run large AI models locally, maintain enormous context windows, generate content, and continuously execute AI workflows on behalf of their users. In other words, Nvidia already put AI in the cloud, but now it’s trying to put AI on your desk. And, possibly, every desktop in every home (running Nvidia software?)

The WiFi theory of AI.

Imagine you’re at home where your AI computer sits somewhere in or attached to your house. Your AI is blazingly fast and doesn’t cost you a nickel because a 120 billion parameter LLM is basically sitting in your basement. You leave home and everything still works, but now you’re using a smaller cloud model with less context and higher costs. It’s less like losing Internet and more like dropping from gigabit fiber to cellular.

Why might that matter? Because the AI sitting in your house isn’t just answering questions. It’s remembering, monitoring, organizing, planning. It’s helping your kids study, tracking travel plans, watching bills, managing subscriptions, updating calendars, and controlling household systems. In short, it’s behaving less like an application and more like your family chief of staff. And your chief of staff doesn’t work one query at a time.

None of these applications require frontier AI. They require persistent AI. And persistence is exactly the sort of thing that gets expensive when every token has to travel through someone else’s data center.

The cable box theory of AI.

So if we agree that there’s this use case (which there is — agree with me, thank you). And we agree that having some sort of AI computer nearby makes this possible (and its usage economically feasible), then who the hell pays for this machine that doesn’t look much different than a $4,500 dev rig?

Will cable MSOs rent you an AI computer for a monthly fee to sign a multiyear contract? Or a mobile phone provider? Maybe.

But what if, say, OpenAI offered “ChatGPT Elite” as a step above ChatGPT Pro. $200/mo for unlimited (specific) model usage. The catch? You need to sign a 2-year contract to get your OpenAI computer that has the biggest OpenAI model that fits in the device. OpenAI gets customers locked in and their data center requirements drop in lockstep with every prompt that can run locally.

OpenAI is already launching some mystery AI hardware device that isn’t a phone, earbud, or anything else anyone has guessed. One friend heard it was a rock. Like, a rock 🪨. And OpenAI CFO Sarah Friar says the device feels “natural” and “lovable.” Which is an interesting way to describe an AI device, and a surprisingly intimate way to describe a piece of hardware. 😳

But, given they shelled out $6.5B to snatch Jony Ive, whose design credits include iMac, iPod, iPhone, iPad, and Apple Watch, maybe it will be a smooth little object quietly running AI all over your house; and, in peak Jony Ive fashion, it will look suspiciously like a very expensive rock. Or, if the job posting tea leaves are any indication, it’s part of a grander hardware ecosystem plan.

What are startups actually doing?

If this thesis is right, I wouldn’t expect the first signal to come from consumer surveys any more than I’d expect technology adoption trends to surface from enterprise IT surveys. I’d expect to see it in startup behavior. Natch.

Startups are already discovering that AI bills can get very large very quickly. (See: tokenmaxxing.) Are startups starting to ask if it would be cheaper to own some of this infrastructure rather than rent it? A Porch Capital canonical thesis is that startup adoption predicts enterprise adoption. So if local AI matters, we should see evidence of it with startups first.

Some things I’d like to measure:

• Mentions of local AI on startup websites.

• Adoption of Ollama and other self-hosted AI frameworks.

• On-device AI architectures.

• Local-first agent platforms.

• Edge AI deployments.

• References to DGX Spark, RTX Spark, and similar systems.

• YC companies building products that assume local inference exists.

I haven’t run this analysis yet, but I’d like to because if thousands of startups quietly start behaving as though local AI is normal, that would tell us that Jensen may not be inventing a new market, but yet again he may be arriving early and making it real. (It wouldn’t be the first or second time.)

Some predictions for 2030. 🔮

Let’s put a few numbers on the board.

What will be the tell — From the Porch.

In 1977, the important question wasn’t whether computers would matter. It was whether every home would eventually get one. (Answer: yep.)

In 1997, the important question wasn’t whether the Internet would matter. It was whether every PC would eventually become an Internet endpoint. (Answer: uh huh.)

In 2026, AI already matters. So the question is will every home eventually become an AI endpoint? Answer: Fuck if I know. But if it happens, I suspect startups will tell us first.

Disclosure: I own NVDA, CRWV, NBIS, and CBRS, and I’m an investor in Groq (SPV) and Crusoe (SPV). Which means I may be completely unbiased, hopelessly compromised, or both. This is not investment advice.

Stage 1: Search.

Everyone starts by using AI like Google.

Fwiw I’m pretty sure my first Google search was “lions” and my first ChatGPT prompt was “tell me about lions.” For some reason, it’s the first thing that pops into my mind when I try out a new platform. No idea why. Anyway, roar / onward.

At this stage, AI is simply a better search engine. A question and answer. Perhaps a little neater because you talk to it more like a human rather than entering search terms, but still minimal (token) consumption.

Stage 2: Synthesis.

Something changes. Instead of looking for a specific piece of information, you start asking AI to synthesize information.

AI stops being a search engine and becomes an analyst – that consumes more AI (tokens) than a stage 1 AI searcher. This is where I think many people first realize AI is fundamentally different from Google.

Stage 3: Creation.

Next comes creation.

At this stage, AI becomes a creator. And (token) consumption starts rising quickly because creation is iterative. Nobody likes the first draft.

Stage 4: Construction.

Then people stop asking AI to create things and start asking it to help build things.

At this stage, AI becomes a collaborator. Users start producing content in addition to consuming it. And they start consuming a lot more AI (tokens).

Stage 5: Delegation.

Finally comes delegation.

At this stage, AI stops behaving like software and starts behaving like an employee. Employees don’t work one query at a time. They work continuously.

The common assumption today is that AI adoption means more users. That’s certainly one way to look at it. But the bigger story may be deeper usage.

A Stage 5 user may consume one hundred times — or one thousand times — more AI (tokens) than a Stage 1 user.

The meter starts running, From the Porch.

The first wave of AI adoption was about getting people to try AI.
The next wave is about moving them from search to delegation.
That’s when the (AI token) meter really starts running.

That’s the entire history of the meeting, in four frames.

I built a version you can drag through — cave to boardroom to whatever 2050+ turns out to be:

Travel Through Time - with The Porch.

Groq is raising up to $650 million (according to Axios) — and not for chips. It (sorta) sold those to Nvidia on Christmas Eve. This raise is for its inference cloud. That distinction is the story.

When Nvidia “bought” Groq, I argued that GroqCloud could be the hidden gem. The market spent five months treating what was left as a fire-sale remnant. Last week, existing investors lined up to recapitalize it pro rata as a going concern, with Disruptive and Infinitum reportedly backstopping the round.

I’m not writing this to take a lap. To be honest, I was wrong (in the rightest way) about the part that scared me. I thought we (Groq investors) got “windsurfed.” We didn’t. But I was right (kinda) about the part that was supposed to be the stretch: GroqCloud as an independent inference cloud.

(Disclosure: I’m an investor in an SPV that holds Groq shares from a 2025 round; since the Nvidia deal, it’s been distributing proceeds (net of fees), so my exposure now is small. The deal’s specifics are confidential and play no part in this piece — everything here is built on public reporting, my own data, and analysis. On the new $650M round: I wasn’t asked to invest, and I know nothing about it that isn’t already in the papers.)

The raise confirms something bigger than Groq: the independent inference cloud — the business of selling tokens, not training runs — has become a fundable category of its own.

Follow the money (into the middle layer).

Look at one window of time:

• Cerebras IPO’d May 14 at $185/share, raised >$5B, and was more than 20x oversubscribed — moving from roughly $8B last September to north of $50B.

• Together AI is reportedly raising around $1B at a $7.5B valuation, on roughly $1B of annualized revenue.

• Fireworks is in talks at a $15B valuation — nearly 4x its October price — on about $315M of ARR.

• Baseten reportedly in talk to raise $1B at an $11B valuation.

• SambaNova pulled $350M+ with Intel co-investing.

• And the latest Groq $650M story.

That’s the market deciding that the independent inference cloud and these six companies — selling tokens, not training runs — comprise a distinct category from the hyperscalers and neoclouds below it and the model labs beside it.

I’ve been writing this thesis for a while, focused a bit more on one layer down the stack. DigitalOcean and Cloudflare and others are a neutral middle (infra) layer — the play against hyperscaler lock-in, the Switzerland between the developer and the giant. Inference clouds like Groq are a similar neutrality bet, moved up a level: serve any open model, fast, cheap, OpenAI-compatible, owned by neither a hyperscaler, nor a neocloud, nor a frontier lab.

What’s actually left at Groq?

Here’s the part worth slowing down on, because the popular read — “they sold everything, the engineers left, there’s nothing there” — isn’t entirely accurate (at least based on what’s been publicly disclosed).

First, the Nvidia deal was a non-exclusive license. Groq kept the LPU IP. They can license it to whoever they want. And the timing is, well, timely given the whole industry is racing toward custom inference silicon to escape Nvidia. The buyers run the stack — Broadcom and Marvell, hyperscalers building their own chips, would-be entrants like Qualcomm, even neoclouds and inference shops like Together sick of renting GPUs. Nvidia paid ~$20B to bring the LPU in-house and, on a non-exclusive license, can’t stop Groq from selling the same blueprint to its rivals.

Second, GroqCloud was explicitly carved out of the Nvidia transaction and kept running. Yes, the founder (Jonathan Ross, godfather of Google’s TPU), the president, and most of the team went to Nvidia. But the chip designs, the fleet already deployed at customers, the developer base, the contracts, a real tokens-per-second speed advantage stayed put. Now, assuming the $650M is real, add to that a pile of fresh cash for interim(ish?) CEO and CFO Adam Winter and Matt Eng to build a next-gen inference cloud.

There’s a cautionary tale, for what it’s worth, that writes itself: Groq raised over a billion dollars, hit a ~$6.9B valuation, shipped a real product to real customers — and still couldn’t reach escape velocity as a standalone in its pre-Nvidia deal form. If that incarnation’s only path was into Nvidia’s arms, what does it say about the cohort now raising at $7B, $11B, $15B…? Or read it the other way: Groq didn’t fail to reach escape velocity — it got there so fast Nvidia had to grab it. Absorption isn’t the consolation prize for a company that couldn’t make it; it’s the receipt for one that mattered.

The demand is real — and bigger than my own data can see.

STAX, the Porch Capital tracker of ~11,000 venture-backed companies’ tech stacks, sees Groq in about 20 of them, Together in 23, Fireworks in 5. Those numbers are tiny, but not because the usage is tiny. Rather, because inference is a server-side call. When a company’s backend hits an inference API, your browser never sees it, so the client-side fingerprinting that nails clouds and CDNs and analytics tools (and the fingerprinting that STAX largely depends upon) is structurally blind to it. The footprint is real. It’s just dark.

For a public read on usage, there’s OpenRouter — the aggregator that’s become the market’s most-watched telemetry, now routing on the order of 100 trillion tokens a month, up roughly 5x in six months. Its providers page publishes routed token volume by named provider — and the board is led by cheap, serve-anything aggregators like Novita, DeepInfra, and SiliconFlow. The premium names you’d expect to dominate (Fireworks, Cerebras, Baseten) sit near the bottom, because OpenRouter is the wrong place to look for them.

But OpenRouter has the opposite blind spot to STAX: it sees only its own channel. Fireworks says it does ~15T tokens per day direct — for 10K+ customers like Cursor, Perplexity, Notion, and Uber — while OpenRouter routes under a tenth of a percent of that. The marketplace rewards cheap, swap-anything providers; the premium players run most of their volume through direct enterprise pipes no public tracker sees. Between STAX’s dark footprint and OpenRouter’s narrow channel, the honest takeaway is that anyone quoting you a single clean “inference market share” number is guessing. The demand isn’t in question. The measurement is.

Inference isn’t just one job.

The conventional wisdom holds that as the world tilts from training to inference, you just cascade old GPUs down to it. Demote last-gen silicon to inference duty, let the new chips train. Cheap, done. That “wisdom” is comfortable and half-right — which is worse than wrong, because half-right stops people looking.

Prefill. It works — to a point. And that point is what the “GPUs are landfill in three years” crowd misses: a depreciated H100 isn’t stranded. Inference has two phases, and the GPU is great at one of them. Prefill — chewing through the input prompt — is compute-bound and parallel, exactly what a GPU, even an old one, does well. As prompts lengthen and agents pack more into context, prefill becomes a large and growing slice of the work. The fleet keeps real value; its job just gets more specific.

Decode. The other phase is where it breaks. Decode — generating output one token at a time — is memory-bandwidth-bound and sequential: every token drags the whole model and KV cache out of memory, so the GPU idles, cores waiting while dollar-per-token leaks away. This is the memory wall I wrote about in The Day AI Runs Out of Memory: inference is a data-movement problem, and HBM is the constraint. The inference-specific chips are built for exactly this — keep more of the model on-chip in SRAM, make fewer off-chip trips, own decode and dodge the HBM crunch in the bargain. It’s also a big part of why Nvidia paid up for Groq’s LPU tech: buying its way into the half of inference its own GPUs are worst at.

Hybrid. So the smart architecture isn’t “one chip does inference.” It’s disaggregation: GPUs for prefill, specialized chips for decode, each matched to the phase it’s good at.

This isn’t theory. In March, AWS and Cerebras announced exactly this — a disaggregated inference service on Amazon Bedrock that splits the pipeline across hardware: Trainium’s dense cores handle prefill, the Cerebras CS-3 handles decode out of on-chip SRAM, and EFA networking ties them together for a claimed ~5x more high-speed capacity. The pattern, named and on the way, from a hyperscaler.

Which reframes the whole accelerator field. Groq, Cerebras, SambaNova et al aren’t Nvidia wannabes chipping at GPU share. They’re the decode-optimized complement to the GPU fleet everyone already owns — which is exactly why the category is forming and getting funded.

And that’s what turns silicon into a business — twice. Almost nobody is going to rack a wafer-scale engine or wire up a Trainium-to-CS-3 handoff themselves; the hardware that makes inference fast is exactly the hardware that’s painful to run. So the winners sell the same chip two ways: the box, to the handful of sovereigns and hyperscalers who can operate it at scale — and the token, to everyone else, as a cloud. Same silicon, two revenue lines. The cloud line may be the one the market is racing to fund next.

The bear case (with an asterisk). 🐻✳️

There’s a real bear case. It’s just smaller than the doomers think — most of it comes with a “but.”

No lock-in. Everyone’s OpenAI-compatible, so switching is a base-URL change; nobody banks a lead, they re-win it every quarter. And the model itself is increasingly fungible — open-weight Chinese models (DeepSeek, Qwen, MiniMax and a wave of newer ones) went from about 15% of OpenRouter’s tracked traffic a year ago to more than half today. You’d think that adds up to commoditization. It doesn’t, quite — when the model is a free, swappable part, the moat just moves down a layer: speed, latency, and reliability genuinely separate the field. The pressure is on rents, not on survival.

Priced for growth, built like infrastructure. Fireworks reportedly raising at ~48x revenue; Cerebras’s 2025 “profit” leaning on a one-off — rich growth multiples on businesses with silicon and datacenters in COGS. But it’s early and not apples-to-apples: a chipmaker and a GPU-renter aren’t the same animal, and today’s margins say little about steady state.

Sovereign concentration. Cerebras drew ~86% of 2025 revenue from two UAE entities; Groq’s marquee bet is Saudi Arabia’s Aramco / HUMAIN buildout — fragile if an anchor walks, and shadowed by export controls (not to mention geopolitical chaos). But that demand is also committed, gigawatt-scale, and multi-year. A shaky revenue base and a funded one at the same time.

The honest read: the bear case is about who keeps the margin, not whether the category survives.

So what is Groq actually worth?

Wrong question — or at least not one anyone answers honestly about private companies whose numbers leak out in fragments. (Also, I don’t know.) The better question is what you’re betting on. And under every one of these companies sits the same bet.

They’re all selling tokens. The price of a token keeps falling — and people buy so many more that total spend climbs anyway. The market grows by getting cheaper, and agentic workloads, firing off dozens of calls per task, are pouring fuel on it. As long as demand outruns price — and right now it’s not close — demand is the easy part.

The hard part is cost, and it’s where they split. Cerebras makes its own chips and runs inference on them. Groq used to; after the Nvidia deal it looks asset-light on silicon — inference on a deployed fleet built around its own IP, without the burden of designing and fabbing the next generation (maybe). Together, Fireworks, and the rest came up the other way, on software — better kernels, better serving, better developer experience — and are now being pulled into owning infrastructure too, because at scale you can’t optimize your way out of your cost of goods.

Three starting points — chip-up, IP-and-fleet, software-down — all converging on owning the infrastructure, for two reasons that compound. The software edge (better kernels, better serving) narrows — the open-source stacks close the gap, and a clever optimization is one everyone replicates, so it stops being enough on its own — even if it’s still (marginally) ahead. And at this scale your economics are your infrastructure bill: rent your GPUs and you don’t control your biggest cost or your supply — especially when your landlord also sells what you sell. So you stop renting and start owning. Hence the datacenters.

So the Groq recap reads as a rational bet on precisely that: tokens become a utility, demand stays bonkers, and a focused inference cloud — freed from the chip treadmill — wins on cost-to-serve. The people who’ve seen the books took their distributions and re-upped pro rata. That’s about the most honest signal private markets give.

The bet was never whether the world wants tokens — it obviously does. It’s who gets to keep selling them, and at what margin, once everyone’s an infrastructure company.

It’s Groqnukkah in June, From the Porch.

On Christmas Eve (when I thought the Jensen stole Groqmas), I said that if GroqCloud became the backbone of token traffic, we’d tell our kids we backed the Maccabees of inference. One scoop later, the miracle oil could last another proverbial night.

And the part I didn’t see coming: it was never really about Groq. That investors are lining up to relight that one candle — if the round closes — shows the room was already full of them — Cerebras and Together and Fireworks each carrying a flame of their own. The leftover was a category all along.

Which candles are still burning in a few years, I can’t call yet. But the room’s lit — and that alone is more than I’d have bet on Christmas Eve.

Happy Groqnukkah. Again.

Disclosure: I hold long positions in $NVDA, $CRWV, and $NBIS. I’m an investor in an SPV that holds Groq shares from a 2025 round; since the Nvidia deal, it’s been distributing proceeds (net of fees). None of this is investment advice — I write From the Porch, not From a Fiduciary Seat.

💜 Short on time? Donate to our Take Steps team here →

We’re past $35,000 for the Crohn’s & Colitis Foundation, and this is the last push before the June 6 walk next weekend. Rachel and I will match the largest individual donation. Most folks give $50-$250; truly anything helps. Everything below is just the why.

Two good things happened in my family this year. I want to tell you about both, because together they’re the entire reason I’m asking you — one more time — to chip in.

Good thing #1: Yesterday, I got scoped. And it was the best report I’ve had in a long time.

If you’ve followed along, you know last fall was rough. A surveillance scare, a removed-but-not-quite-resolved lesion, and a December procedure (an endoscopic mucosal resection) that finally took the surgery conversation off the table — for now. After all that, I went in this week bracing for the worst.

Instead, given my history (decades of ulcerative colitis, prior high-grade dysplasia), the report was about as good as I let myself hope for:

• Mayo score: 0. No active disease.

• Bowel prep: 9 out of 9. A perfect score. (I will be adding this to my LinkedIn.)

• Every segment normal: top to bottom.

They removed three tiny things and took a lot of biopsies — including, very deliberately, the spots that have scared me before. The honest caveat: pathology is still pending. The tissue is what actually answers the question, and after the year I had, my brain doesn’t get to declare victory until the lab does. But the picture today is a good one, and I’m going to let myself have it.

Bonus content: Scope Prep Pro Tips.

∙ Sutab pills > liquid preps. And FFS avoid GoLYTELY ← it comes in a gallon jug.
∙ Gatorade + water ← it’s the oz of fluid, not the dozens of pills. Gatorade is >92% water.
∙ Propofol > fentanyl. Best sleep ever (anesthesiologist required); or wake up and barf.
∙ Use an AI chatbot for the journey. I use ChatGPT for this to save on costs (and dignity). It’s genuinely helpful, will answer your questions and keep you sane, and may give you gems like this: “I have consumed enough water to irrigate a golf course and my colon is producing sports-drink scented weather patterns.” 😆🚽🥤” ←Real.

(If you’re new here and wondering why a guy writes this much about his own colon: I bought a camera for my toilet, I own LogMyLog.com, and this is simply who I am. Welcome.)

Good thing #2: Soleil had her first winter without a flare.

My daughter was diagnosed with Crohn’s at six. If you met her today, you’d see a horse-obsessed 10-year-old with bottomless energy. What you wouldn’t see is everything the disease has already put her through — the hospitalizations, the procedures, the medication roulette, the infusion reaction that turned her lips blue mid-IV.

This past winter was different. For the first time since she was diagnosed: no middle-of-the-night ER sprints, terrifying weight loss, or medication roulette. She skied. She rode horses. She had sleepovers. She got to be a kid.

🎥 Watch: Soleil’s Warriors 2026 — two minutes, in her own words. It’ll do more than anything I can write here.

Neither of these was luck.

I still have a colon because of timing and tools — chromoendoscopy, AI-assisted detection that once flagged a 2mm high grade dysplasia a tired human eye might have missed, doctors who re-biopsy the scary spots on purpose. Five years ago, my chart said colectomy. This week it said “see you in a while.”

And Soleil’s quiet winter happened because the medication keeping her flare-free exists. The biologic that put her into her first remission exists. So does mine.

That’s not a metaphor. It’s the literal dividend of research — the kind the Crohn’s & Colitis Foundation has funded at every major step for decades.

So, the ask. (Same one. Last time before June 6, I promise.)

I know I’ve asked a lot of you. Emails, texts, posts, the (more than) occasional guilt trip. You’ve shown up — last year our team raised over $35,000 and landed in the Top 10 nationally. That genuinely floors me.

This is the final push before the walk. And because people would rather join something that’s already working than start from zero, let me just say it plainly:

We’re past $35K. Help us finish strong.

👉 Donate to our Take Steps team here. Rachel and I will match the largest individual donation. $50, $250, $1,000 — whatever feels right. It funds better treatments today and cures tomorrow: for Soleil, for me, and for the 10 million people living with IBD who never got to choose it.

One confession: I won’t actually be at the walk — I’ll be in New Jersey with my brother’s family. But Rachel and Soleil will be there at North Point Park in Cambridge, and they’re in good hands: Rachel is on the Foundation’s New England board and is co-chairing Take Steps this year. So I’ll be there the only way that actually counts — through whatever we raise together before June 6th.

With love and gratitude (and a perfect 9/9),

David (and Rachel, Soleil, and Aden).

I am not the best at anything.

If you’re looking for the world’s foremost expert on AI infrastructure or venture investing or founding companies or trading stocks or any one thing, I’m probably not your guy. There are smarter people on every one of those topics.

What I am is sneaky-dangerous across a bunch of them simultaneously.

That’s the decathlete. Not the fastest, not the strongest, not the best jumper. But when you add up all ten events — when you need someone who can run, jump, throw, and think — the decathlete wins the whole thing.

I spent years being (more than) slightly embarrassed by this. The guy who could hold his own in almost any room, but never fully belonged in one. The investor who was also a founder who was also an operator who was also an analyst. Pick a lane, kid.

I never did. Turns out, I was trained not to.

At Penn I did the M&T dual degree program across Wharton and engineering. I suppose it was a template for what was to come.

The events, roughly in order:

• 1996-2001. Sell-side research and investment banking. Furman Selz / ING Barings, then (JPMorgan) Chase H&Q covering Internet infrastructure. Financed the dotcom boom (bubble), watched it explode (burst), learned what hubris at scale actually looks like.

• 2002-2005. Hedge fund partner and analyst. MLH Capital with Ed Hajim and then Ardsley Partners. Ed taught me two things I’ve never forgotten: hubris is terminal, and the long term is just a string of short terms put together. Then he fired me. (And got me the job at Ardsley, who also later fired me.) I’m a better analyst now than I was then — experience gives you pattern recognition and epistemic humility simultaneously. (Whether you wanted it or not.)

• 2007-2010. Co-founded Tigerbow — send a real gift via email. Built it from concept to live platform. Wound it down. That’s the polite version.

• 2010-2011. Co-founded Philo — Foursquare for TV. Raised >$1M, built the product, negotiated an acquisition to a private company that later failed too. I made no money. What I got was “the education.”

• 2012-2014. EIR at Broadway Video Ventures (Lorne Michaels’s thing) and Comcast Ventures. Sourced deals, explored ideas, and — according to my dear friend David who ran Broadway Video Ventures — spent a meaningful amount of time on his couch working on my mortgage refi. (Not entirely untrue.)

• 2014-2018. AWS, first tour of duty. Co-built the startup BD team — the playbook, infrastructure, startup GTM effort. Helped expand coverage to >100K startups and >10K partners. Learned what platform-scale institution building actually feels like.

• 2018-2019. Stripe. Left AWS to help co-build the startup growth team at one of the great hypergrowth companies of the era. Stripe was — and is — a remarkable company. The role was… not what I expected. I left after a year and a day. (Not a coincidence.)

• 2019-2023. AWS, second stint. At Amazon they call this a boomerang. I call it Hotel California — you can check out any time you like, but you can never leave. (Until they do another dozen RIFs.) Was nice being back, but no longer felt like I was in the room where it happened. But it was home until it wasn’t.

• 2023-present. Porch Capital. Which brings us here.

Throughout all of it, angel investing — in CoreWeave, Twitter and LinkedIn pre-IPO, Groq (pre-Nvidia), Crusoe, Throne Science, and others. Not as a strategy, but as a compulsion. Or maybe an “entrance fee.” It’s an expensive hobby. Being a great angel investor is a wonderful life, but you can only get there by being ultra-successful or being born into money. VCs are paid to invest. An angel’s job is to be rich. (It’s a very nice gig if you can get it.)

I am neither, which makes me either an idiot or an optimist. Or both.

None of these events, alone, makes me the best in the world at anything.

All of them together make me someone who can look at an AI infrastructure company and think about it from the perspective of a founder, operator, investor, analyst, and builder — simultaneously, in real time, without having to go look things up. (Zip it, ChatGPT and Claude — and sometimes Gemini 🤫.)

That’s what Porch Capital is. Not a fund. Not a firm. Not (just) a newsletter. A decathlete sitting on the porch with something to say.

I spent a long time thinking I should pick a lane. I’m done with that….

Porchdog. Decathlon Champion.

What can I say. I owe it all to Little Chocolate Donuts. (RIP, John Belushi.)

Those mammoth deals sucked the air out of the AI echo chamber just as three separate organizations quietly built the exchange infrastructure that will determine how the world prices AI.

• CME Group partnered with Silicon Data to create a futures market for computing power.

• The NYSE’s parent company Intercontinental Exchange announced GPU compute futures contracts with a pricing data firm called Ornn.

• Brett Harrison, former president of FTX US, already went live with compute futures on his derivatives exchange (Architect), tracking H100 and H200 rental prices in real time.

Larry Fink called compute a new asset class and H100 secondary prices jumped almost 50% since the end of April.

Fwiw, I wrote all of this eight weeks ago. As satire.

The fictional CME is now just the CME.

In March I published When Compute Becomes a Financial Asset. The thesis: AI infrastructure is inventing a new financial system and Nvidia is the central bank. I ended it with a fake 2036 news dispatch from the “Compute Mercantile Exchange” and the “Huang Monetary Authority.” I thought I was being clever. Evidently I was being early.

I’m not saying I predicted this. (Well, maybe.) I’m saying the transition is happening so fast that satire has a shorter shelf life than a GPU generation. (Either way, I’m taking the W.) That speed is itself the signal. The CME isn’t just speculating on the latest fad asset class; it’s responding to demand that already exists. The question: “Is compute becoming a financial asset?” is moot. Now, the question is why does this market need to exist today? And (also) what does it unlock?

The jet fuel problem.

Here’s the thing people are more than whispering:

Capital is moving from labor to compute.

Consequently, compute is becoming the primary cost of doing business for an entire generation of companies. But unlike salaries — which are fixed and predictable and budgetable — compute pricing is volatile in a way that makes CFOs shit blood and drink at breakfast.

Remember, H100 spot prices jumped 50% in a month. That’s an existential swing if you’re an AI-native startup. And it will be an existential nightmare for tomorrow’s Fortune 500 companies that see their capital allocation to compute grow far greater than labor. You can’t plan a business around a cost input that moves like that. You can’t raise a Series B with a cost structure that volatile. You can’t sign an enterprise contract at a fixed price if your underlying (compute) cost is a moving target.

A dash of perspective: the jet fuel market is ~$200-300B annually depending on the source. But the futures market for jet fuel is 10-30x that because of leverage and speculation. The airline industry figured out decades ago that you can’t run a business against a cost input that volatile without a futures market.

Now a dollop of context: Big Tech alone is projected to spend $655B on AI infrastructure in 2026. ​And ​Don Wilson (founder of DRW and Compute Exchange and backer of Silicon Data) said it plainly: “Compute will become the largest commodity in the world.” So a futures market on compute could be 10-30x the size of an industry twice as large as jet fuel and growing faster than a jet plane.

Every startup has this problem…

…And most don’t know it yet. Across the 11K+ venture-backed companies tracked by STAX (Porch Capital’s proprietary startup signal monitor), almost 2K are emitting a detectable AI signal. Of those, ~93% are running OpenAI. Of the almost 250+ companies with a detectable AI dev tool signal, 68% are running Claude Code and 56% are using OpenAI’s codex. That’s panic-inducing concentration: two vendors, pricing players, and supply dependencies rippling across an entire generation of startups (and soon enterprises) simultaneously. Every one of those companies has compute cost exposure that they currently have no tool to hedge.

This is the jet fuel problem. Fuel is the airlines’ biggest variable cost and fuel prices swing wildly with oil markets. So airlines buy futures. They lock in their cost exposure months in advance. It doesn’t eliminate the risk, but it makes it manageable, predictable, and financeable.

AI companies now have the same problem with compute that airlines have with jet fuel. The CME just built the hedging instrument they’ve been missing.

The financing unlock is bigger than hedging.

The structural financial problem underneath all AI infra (and consequently all layers of the AI stack) is that traditional lenders — banks, private credit funds, infrastructure investors — want to finance AI compute infrastructure given the demand and capex requirements, but they can’t underwrite an asset they can’t price. You can’t write a loan against a cluster of H100s if you don’t know what an H100 hour is worth in six months. There hasn’t been a forward curve, benchmark, or index.

This is what I wrote about in March. Nvidia has essentially been forced to finance its own customers because nobody else could or would. The circular logic was always going to be a problem: Nvidia funds CoreWeave, CoreWeave buys Nvidia GPUs, Nvidia books the revenue. Michael Burry and a handful of bearish analysts throw a hissy fit. But it kept happening because there was no alternative. Capital couldn’t flow into AI infrastructure at scale otherwise because the asset couldn’t be priced at scale. (Well, plenty of capital flowed, but it wasn’t from traditional banks to AI models and infra.)

A compute derivatives market mitigates that. Once CME and ICE establish a benchmark index — a real, tradeable, regulatorily-reviewed price for a unit of compute — suddenly a bank can write a loan against a neocloud’s cluster the same way they’d finance an oil storage facility or a power plant; i.e., against a known forward curve, with real collateral, and with a real hedge.

The derivatives market is the missing piece of plumbing that greases the rest of the financing wheels. Blackstone can deploy $5B into a Google TPU JV partly because there’s now a credible path to pricing the commodity inside the container they’re building.

The CME was hardly first.

The CME and ICE didn’t invent this idea, but they did validate it. A shtickle of startups has been striving to spin up and scale compute markets for several years — many some combination of ignored, underfunded, or too early. Sound familiar? (Yes.) It’s how every commodity market starts.

Ornn raised a $5.7M seed round in October 2025 explicitly to build compute futures infrastructure. Their Compute Price Index — built from actual printed transactions, not surveys or broker quotes — landed on Bloomberg terminals in April 2026. The first compute swap cleared in December 2025. Six months before CME called this a “multi-trillion-dollar market opportunity,” Ornn was already clearing trades. Now they’re the index provider behind ICE’s futures contracts. That’s the startup playbook working exactly as intended.

Compute Exchange — co-founded by Don Wilson of DRW and Simeon Bochev, formerly of Apple and Lambda Labs — was running spot auctions for GPU compute since early 2025, facilitating over $1B in compute supply. Their data showed companies being overcharged by as much as 40% because there was no price transparency. A futures market fixes exactly that.

SF Compute raised a $40M Series A in December 2025 at a $300M valuation — backed across rounds by DCVC, Wing, and Factorial Capital (disclosure: I’m a Factorial Capital LP, making it all the more embarrassing that I’m only writing this now) — building what it calls the Airbnb of GPU capacity. A two-sided marketplace with buyers reselling unused capacity; no GPU ownership required. They now manage over $100M in hardware without owning a single chip.

Andromeda started as a GPU procurement operation inside NFDG — the near mythic venture fund run by Nat Friedman and Daniel Gross. When Zuckerberg hired both of them for Meta Superintelligence Labs in mid-2025 (following a separate $1B+ tender offer to buy out their venture fund LPs), they spun Andromeda out as a standalone company under CEO Wil Moushey. It raised $60M from Paradigm at a $1.5B valuation in March 2026, hit a $100M revenue run rate in 2025 — profitably — and now connects AI teams to 100+ compute providers with standardized contracts. Moushey’s line: “Our competition is unutilized GPUs.”

And then there’s Barkr — doing something completely different and, possibly, more important. While others are building trading infrastructure, Barkr is building the valuation infrastructure. They provide insurance-backed appraisals of GPU clusters so lenders can actually write loans against them — and if they get the price wrong, they pay the difference. In the past six months they’ve touched about $200M in GPU transactions. Nvidia is quietly sending them lenders. Their CEO Thomas Galbraith put it best: “We make them boring assets.” That’s the whole game. You can’t finance what you can’t price, and you can’t price what no one will insure.

The pattern across all of these: startups identified the problem, built the infrastructure, established the indexes and benchmarks. Then CME and ICE looked at the data and decided the market was real enough to regulate. That’s how commodity markets form. The wildcatters go first; and the exchanges follow.

Read the job postings. (You may be a fit.)

Companies telegraph their strategy in hiring before they announce anything publicly. They make for interesting reading if you’re into this sort of thing.

OpenAI posted a role for Head of Compute Capital Markets. The JD says:

• “…translate compute demand into investable frameworks for lenders and investors….”

• “…build durable financing architecture for hyperscale AI infrastructure….”

• “…develop innovative financing structures across debt, structured capital, and alternative financing vehicles.”

That’s a capital markets banker. Inside an AI lab. Sitting next to the people building GPT-6.

Anthropic — the company whose AI helped me research and draft this piece, which I find both useful and mildly unsettling — posted JDs for at least three finance roles that didn’t exist two years ago:

• Head of Capital Markets & Corporate Development.

• Finance & Strategy, Compute Infrastructure lead.

• Capital Markets & Investor Relations.

Together AI hired a DCM executive — the kind that used to sit at a few big banks. Together shows up in the inference stacks of startups tracked by STAX — their DCM hire isn’t random, and it isn’t about models.

xAI — that little AI model lab inside that soon-to-be public rocket company — posted roles for experts in structured finance, private credit, and DCM investment banking. Ostensibly to train their AI models on financial concepts. Maybe. Or maybe you hire that many structured finance people because you’re about to structure something.

Mistral, meanwhile, skipped the hiring and just launched a compute product. European-hosted GPU capacity, bare metal to managed clusters, sovereign data center certifications. They went down the stack entirely rather than hire bankers to finance the one they’re on.

The pattern across all of them is the same: every major AI lab is simultaneously building financial infrastructure around compute costs. Treasury functions. Debt capacity. Infrastructure finance. Hedging frameworks. They’re all doing it because they all have the same problem — compute is now the largest, most volatile cost input in their business, and none of them have the tools to manage it properly.

All of which raises the obvious question: who trades this thing?

Who’s long and who’s short?

AI-natives are natural buyers. They want to lock in compute costs the same way a manufacturer locks in steel prices: predictability over optimality. The inference layer alone — where startups and AI-natives actually burn compute — is currently split across 14+ providers in the STAX data, from Modal (another Factorial Capital portfolio company, fyi) to HuggingFace to Together, each pricing opaquely with no shared benchmark. (Caveat: STAX monitors directional signal; it is not a census.) That fragmentation is exactly the problem CME is trying to mitigate. You can’t hedge what you can’t price.

Hyperscalers are natural sellers — not because they have excess capacity (they don’t; every earnings call this year has been about not having enough). But because locking in forward revenue at known prices is strategically valuable when you’re committing hundreds of billions in capex against uncertain future demand; and when you’re reporting all this to trigger-happy investors once a quarter. AWS, Google, and Azure would love to sell futures contracts the same way a power plant sells forward electricity — build it, price it, move on.

Neoclouds are the most interesting and potentially most exposed. CoreWeave and its peers are simultaneously long physical compute — they own the clusters — and exposed to spot price drops if compute gets cheaper. A futures market cuts both ways for them. It enables hedging, sure. But it also creates transparent price discovery that makes it much harder to extract margin in what has until now been a gloriously opaque market. Watch how they respond to CME and ICE going live. The silence will be telling.

And then there’s the speculator class — the hedge funds, quant desks, commodity trading firms that will sniff out the arb between spot and futures pricing the moment there’s enough liquidity. Don Wilson didn’t just say compute would become the world’s largest commodity. He went and built the exchange to trade it. That’s conviction from someone who made billions in energy derivatives — and it should tell you everything about who’s going to show up when these markets go live.

From the Porch: the compute exchange is already open.

Blackstone got the headlines last week for writing a $5B check to build TPU data centers with Google. It’s a big deal — PE money flowing into AI infrastructure at this scale is real and important. But Blackstone is financing the container: the building, power, the chips, and physical infrastructure. That’s the old model in a new(ish) wrapper.

CME just priced the commodity inside the container. That’s the new model. And historically, when commodity markets form around a new input — oil, electricity, bandwidth — the financial infrastructure built around pricing that commodity ends up being at least as valuable as the physical infrastructure built around producing it.

The compute exchange is already open. Most people don’t really know it exists yet.
That’s usually when it’s worth paying attention.

Disclosures: long NVDA, CRWV, NBIS. Equity positions in Groq (mostly “acquired” by Nvidia, 2025) and Crusoe via SPVs. LP in Factorial Capital, whose portfolio includes SF Compute and Modal. STAX data referenced above covers 11K+ venture-backed companies as of May 24, 2026 — directional signal, not a formal dataset. Nothing here is investment advice — I’m just a guy yelling about stuff From the Porch.

PS: subscribe if someone forwarded this to you.

None of that happened.

Graphcore sold to SoftBank in July 2024 for less than what it had raised — about $500M against $767M in total funding and a peak valuation of $2.8B. The co-founders are still there. The chip still exists. And SoftBank injected another $457M last week, which you also probably didn’t read about.

The official cause of (near) death — depending on how you score a SoftBank acqui-rescue — was (ostensibly) Graphcore’s own software. A former ML engineer told Sifted: “The software was just not ready. There were lots of bugs and issues.” Microsoft walked. Investors wrote down their stakes. By late 2023, Graphcore was close to running out of money.

But there’s another version of the story.

The version that circulates in AI infrastructure sewing circles goes like this: Microsoft was ready to deploy Graphcore IPUs at scale. The deal was real. And then Nvidia — which at the time tightly controlled its (very limited) supply of its H100s, with waitlists stretching 12 months and prices to match — made clear that being seen with a competitor had consequences. The Microsoft deal didn’t happen. Graphcore ran short on runway. The chip that benchmarked competitively against the A100 on inference workloads died not in a performance test but in a procurement meeting.

Is that true? Hell if I know. What is on the record, however, is (former) Groq CEO (and now Nvidia “Chief Software Architect”) Jonathan Ross, speaking to the Wall Street Journal: “A lot of people that we meet with say that if Nvidia were to hear that we were meeting, they would disavow it. The problem is you have to pay Nvidia a year in advance, and you may get your hardware in a year, or it may take longer, and it’s, ‘Aw shucks, you’re buying from someone else, and I guess it’s going to take a little longer.’”

Scott Herkelman — a former head of AMD’s GPU division (and also former Nvidia executive) — backed it up saying: “This happens more than you expect. They learned from GPP to not put it into writing.”

So the mechanism is documented. Whether it killed Graphcore specifically is the urban legend part. But the urban legend exists because the mechanism is real. And in 2022, when H100s were the scarcest resource in technology, the mechanism had teeth.

Something shifted. And you can date it pretty precisely.

The H100 shortage didn’t end, but it did loosen. TSMC ramped. Blackwell arrived and reshuffled the queue. The waitlists that once stretched a year began to move, even if the price didn’t drop and availability never got easy. And with that loosening, however partial, went some of Nvidia’s most powerful non-technical weapon — the ability to make the cost of choosing differently existential. You can’t threaten the back of a line that’s gotten shorter. You can still threaten the back of a line, just with less teeth.

But the deeper shift was the workload (not just chip availability).

Inference grew up. Training — where Nvidia’s dominance was built and where CUDA’s switching costs are most brutal — is increasingly a known quantity. Inference is where AI actually touches users, generates revenue, and compounds. And inference, it turns out, has different physics than training. Inference increasingly becomes a memory-bandwidth problem. Every generated token forces enormous model-weight and KV-cache movement through memory. Do that billions of times a day across agentic systems that call each other recursively, and the GPU — designed for parallel compute, not memory throughput — starts to show its limits.

This is exactly what Cerebras built for. The WSE-3 pushes 21 petabytes per second of on-chip memory bandwidth. On raw on-chip memory bandwidth, Nvidia’s Blackwell isn’t even in the same conversation. Storing the entire model in on-chip SRAM means no memory wall. No fetching. Just tokens, fast, at a scale that changes what’s possible for real-time and agentic workloads.

AWS saw enough to bet on this. And AWS — the most vertically integrated, self-sufficient cloud on earth, with its own inference chip (Inferentia), its own training chip (Trainium), and its own semiconductor division (Annapurna Labs) partnered with Cerebras in March to deploy a disaggregated inference architecture across Amazon Bedrock: Trainium handles prefill, the compute-intensive phase of processing your input. Cerebras CS-3 handles decode, the memory-bandwidth-intensive phase of generating your output. Two chips, two architectures, one inference pipeline — because no single chip does both phases optimally.

AWS didn’t do this because Cerebras was cheaper. It didn’t do this because Nvidia failed. It did this because the inference workload finally got weird enough that physics won. And when the most self-sufficient buyer in the world goes outside, that’s a tell.

AWS wasn’t alone. OpenAI — the company that has done more than anyone to make Nvidia indispensable — signed a $20B Master Relationship Agreement with Cerebras for 750 MW of inference compute, expandable to 2 GW. OpenAI also loaned Cerebras $1B secured by warrants. Sam Altman is an investor, and OpenAI’s infrastructure chief is quoted in the S-1. At some point the question isn’t why companies are choosing Cerebras. It’s why that surprises anyone.

(Oh — and the Annapurna executives who built Trainium? Two senior leaders left within seven months of each other. Amazon said nothing.)

Here’s the part nobody is writing about: The Silion Trinity.

While the industry has spent three years watching AMD (attempt to) chip away at Nvidia’s margins, and hyperscalers build custom silicon for their own internal workloads, Softbank has been quietly assembling a replacement stack. Masayoshi Son even gave it a name. Project Izanagi — after the Japanese god of creation. A $100 billion venture with the asset base already in place:

• Arm Holdings: the architectural layer — until March 2026, when they announced their own chip. For 35 years Arm only licensed its architecture. Then they built the AGI CPU with 136 cores, co-developed with Meta, and with OpenAI among its first customers. The blueprint company became a chip company overnight. Softbank acquired it in 2016 for $32B, held it through a decade of skepticism, and relisted in 2023. Arm’s architecture now powers virtually everything at the edge and increasingly in the data center — including, in a delicious irony, Nvidia’s own Grace Blackwell superchips. Nvidia builds on Arm. Softbank knows this. Now Arm competes with Nvidia directly for control of the AI server stack.

• Ampere Computing: the CPU layer. SoftBank acquired the company (founded by Renée James, former president of Intel) in November 2025 for $6.5B. Arm-based server processors with up to 192 cores, designed from the ground up for data center AI workloads. Energy efficient. Cloud native. Not x86.

• Graphcore: the accelerator layer. Acquired in 2024 in a near firesale plus $457 million injected last week. The IPU that couldn’t survive Nvidia’s leverage as an independent company is now the AI accelerator inside Softbank’s new stack — paired with Ampere CPUs and built on Arm architecture, integrated at the silicon level with 2.8 terabytes per second of interconnect bandwidth.

“They” call it the Silicon Trinity. Arm at the foundation. Ampere as the brawn. Graphcore as the intelligence layer. A closed-loop system designed from the ground up — not an x86 CPU bolted to an Nvidia GPU, but a co-designed stack where every layer speaks the same architectural language.

This is not AMD, which is fighting Nvidia on Nvidia’s terms — same GPU architecture, same workload paradigm, and hoping ROCm gets good enough (so far: nah). SoftBank is doing something categorically different: building the first credible full-stack alternative to the x86 + Nvidia pairing that currently runs the entire AI industry.

The Graphcore that couldn’t IPO in 2024 because it ran out of money — and possibly because Nvidia played a little dirty — is the same chip that’s now at the center of the most ambitious anti-Nvidia project on the planet. In the alternate history, it IPO’d and became a public company. In this one, it nearly died, got rescued, and became a weapon.

Masayoshi Son has been spectacularly wrong before: WeWork, Sprint, The Vision Fund. But he has also been right in ways that took a decade to see — and Arm, his most patient bet, is now the substrate everything runs on, including his supposed enemy.

Look around, From the Porch.

Few are watching SoftBank’s chip stack. That’s exactly when you should be. The inference race is open. Cerebras proved it on day one of its IPO. Graphcore is proving it in a different way, on a longer timeline, inside a stack most people haven’t looked at closely enough.

The chip was never the problem. It was always the leverage. And leverage, it turns out, has a shelf life.

Disclosures: I am currently long $NVDA, $CRWV, $NBIS; I have equity in Groq via SPV (mostly “acquired” by Nvidia, 2025). No position in Cerebras, Arm, or SoftBank. Not investment advice — I’m an analyst that writes, not the other way around.

Full text of February 14, 2014 USAi Tomorrow newspaper article for those of you that can’t read 4 point text….

Graphcore, the Bristol-based maker of Intelligence Processing Units, priced its long-anticipated Nasdaq IPO at $18 per share Tuesday, raising $1.2 billion in the largest AI chip debut in recent memory. Shares closed at $30.24, up 68%, as investors bet the company’s purpose-built IPU architecture and marquee Microsoft Azure partnership signal the first credible crack in Nvidia’s GPU dominance.

“Today validates what we’ve believed since 2012,” said Nigel Toon, Graphcore’s co-founder and CEO, speaking from the Nasdaq MarketSite. “AI doesn’t need a repurposed gaming chip. It needs a processor designed from the ground up for how intelligence actually works.”

The IPO follows a pivotal 18 months in which Microsoft’s Azure platform deployed Graphcore IPU clusters at scale, providing the enterprise proof point that had eluded the company for years. Azure customers running inference workloads on IPU hardware reported latency reductions of up to 40% compared to equivalent GPU configurations.

The Race is Now Open

Analysts called the debut a turning point for a market long assumed to have a single winner. “What Graphcore proved is that the inference workload has different physics than training,” said one semiconductor analyst. “The GPU was never the only answer.”

The IPO prospectus disclosed $340M in revenue for FY2023, up 8x y/y, driven almost entirely by the Azure deployment. Gross margins of 61% exceeded expectations.

Nvidia: “Competition Is Healthy”

Jensen Huang, Nvidia’s chief executive, issued a brief statement Tuesday afternoon as shares of the company fell 4.2% following Graphcore’s debut. “We welcome competition,” Huang said. “It makes everyone better.” Nvidia did not comment on whether the company had delayed GPU shipments to customers in discussions with Graphcore.

Analysts were less measured. “This is the first time in years a pure-play AI chip alternative has had a public market validation,” said one semiconductor analyst who covers both companies. “Nvidia’s moat is real. But moats get crossed.”

What Comes Next

Graphcore’s IPU-M — its next-generation wafer-scale chip — is expected to enter production later this year, according to the prospectus. The company said it has letters of intent from two additional hyperscalers it declined to name, beyond the existing Azure relationship.

Analysts set a price target of $38.

But none of it changed the thesis.

TLDR: the world’s computing infrastructure has to be rebuilt (not just upgraded) because AI compresses the old stack. The traditional three-tier architecture (front-end, logic, database) is collapsing into a single layer. The LLM is the logic and the database and (increasingly) the front-end. And that thing runs on GPUs. Every server rack, data center, edge node built for the old world is the wrong tool. The bill to rebuild it is $2T+ in compute revenue before it plays out. And it’s not all going to AWS, GCP, and Azure — which is why the neocloud thesis exists, and why I’ve been writing about it since 2024.

Last week, the market started grading the rebuild era more broadly. It was looking for proof. Not just quarterly results, AI exposure or proximity, or new taglines on websites announcing AI branding. Here’s what the report cards looked like — and what to do.

The Honor Roll.

DigitalOcean (DOCN) beat and raised again (again). Revenue was $258M (+22%) and AI customer ARR (+221%) is now 17% of the total mix. They guided for 26% growth this year and, perhaps audaciously, 50%+ next year. The stock gapped up 30%+ the next morning.

I’ve been writing about DOCN’s arc for a while now. Here’s the capsule version: three years ago it was a second- (third-?) rate cloud provider losing customers up and down the stack. They bought Paperspace to get into AI in 2023, but their existing business subsequently fell apart. It was ugly: they missed a quarter, restated historicals, lowered guidance, and fired the CEO (and got “rewarded” with a ~50% stock decline to the $20-range). New management, new positioning, three consecutive beat-and-raise quarters later — the numbers are catching up to its (new) “AI Native Cloud” branding. AI ARR is real, it’s growing, and now it’s large enough to move the top line. The market loves its AI story and (even more) its revenue growth acceleration forecast.

This is exactly the middle-layer story to which I keep returning. DOCN is running models. That’s the business.

Nebius (NBIS) popped on the $643M acquisition of Eigen AI — an inference and model-optimization company — which I wrote about in detail here. 1Q26 earnings don’t drop until May 13, so this is a pre-earnings pop on deal conviction, not a numbers reaction. (Translation: you’re not done nail-biting.) The market is buying the map: GPU infrastructure + optimization + inference deployment, all in one place.

The Eigen deal fills the exact gap that sits between “we have GPUs” and “we make a product.” This is where NBIS just planted its flag.

The Paradox.

CoreWeave (CRWV) had a genuinely strong quarter. Revenue $2.08B (+112% y/y) beat estimates. Revenue backlog approached $100B (from ~$66B). Ten customers committed to spending more than $1B each. A $21B deal with Meta and a multi-year deal with Anthropic signed in the quarter. “We have reached hyperscale,” the CEO said.

Aaaaand… the stock slid 10%.

This is the third consecutive “beat-and-punish quarter.” It happened in 3Q25 and in 4Q25, and it happened again. The market isn’t doubting the revenue. It’s worried about $25B in debt (notably supported by long-term contracts from investment grade buyers), $31-35B in capex guidance, and an EPS miss of -$1.40 versus -$0.92 expected. And the market evidently concluded that CoreWeave’s path to profitability remains a show-me story.

My view hasn’t changed: CRWV is the undisputed leader in AI infrastructure. Its capital structure, to be fair, is legit anxiety provoking. But both things can be true. If you believe in the GPU cloud thesis — and I do — this is the dip. I’m buying.

The Weird One.

Cloudflare (NET) is the most perplexing of the week’s results. And possibly the most misread.

Revenue was $640M (+34% y/y) and management raised full-year earnings guidance, though 2Q26 guidance was a bit light. They’ve got Workers AI, AI Gateway, Vectorize, etc. – a real product roadmap for the agentic AI era. The stock had been up 30% YTD heading into the print. But with the earnings release they announced a 20% workforce reduction (1,100 people) framed explicitly (and ironically af) as a shift to an “agentic AI-first operating model.”

Aaaaand… yikes. The stock dropped ~24% ($60!) on Friday.

The irony, of course, is that the best example of the bull case for NET is in the layoff announcement. If AI enables them to operate at 20% lower headcount while growing revenue at 30%+, the long-term margin story gets materially better. That should be the signal. Instead, the market saw restructuring charges, a barely-OK guide, and a stock with no room for ambiguity. So, investors dumped (or, more descriptively, vomited) the stock.

From what STAX tracks across 10K+ VC-backed companies: Cloudflare appears in the stack more consistently than almost any other CDN or edge provider. Workers, R2, WAF — it’s everywhere. The underlying adoption is real. And this all looks like an overreaction.

The Legacy Pivot Race.

Akamai (AKAM) reported in-line: $1.07B revenue (+6% y/y), EPS at expectations. Nothing exciting. Then they dropped a $1.8B, seven-year commitment from an unnamed US-based frontier model provider for cloud infrastructure services. Stock up ~26%.

This is the market paying a very high premium for proof of concept — validation that AKAM can land an AI anchor customer at all. The frontier lab – unnamed, presumably (and hopefully?) inference / deployment workloads – is essentially vouching for Akamai’s distributed infrastructure as a viable AI substrate. Revenue doesn’t even start until 4Q26, but the signal mattered: a legacy CDN just got picked by a frontier lab. That’s a different stock than it was for most of the rest of the year during which AKAM lagged its group.

Whether the rest of the business can execute to match the implied new valuation is a separate question.

Fastly (FSLY) beat on revenue ($173M, +20% y/y), crushed EPS ($0.13 vs $0.09 expectation), and raised full-year guidance.

Aaaaand… the stock plummeted 38%.

Keep in mind FSLY had run +441% over the prior year on the hope of a DOCN-style AI transformation. The market had priced in a pivot — AI-native product, inference-layer revenue, a new reason to exist in the stack. What they got instead was a well-run CDN quarter. Solid fundamentals. 11% growth in a core business. But not so much AI-specific product gravity. No 221% AI ARR growth. No frontier lab deal.

Fastly is doing fine. It’s just not doing the thing the stock implied it would do. The gap between narrative and reality closed violently.

The Thread.

Six stocks. Most beat. One pattern.

The market stopped rewarding AI adjacency and started grading on AI revenue conversion. How complete is your pivot? How real are the receipts? How much rope do you have while we wait?

• DOCN: conversion proven, three beats in a row, receipts real → rewarded. ✅

• NBIS: (re)born AI, building the inference stack, map credible → rewarded. ✅

• CRWV: best AI revenue in the group, punished for debt etc. → punished (buy the dip). 🤦‍♂️

• NET: executing well, confused the market with the RIF → punished (overreaction?) 🤮

• AKAM: one big contract, partial credit, execution TBD → rewarded (on faith). ✅

• FSLY: ran on the story before writing it → punished (i.e., got the bill). 📉

From STAX: across 10K+ VC-backed companies, OpenAI/GPT appears in 10% of stacks (766 companies), Anthropic/Claude in 5% (363). The detectable inference layer — Modal, vLLM, Replicate, Baseten, Cerebras — is still only ~60 companies in total. Self-hosted models don’t leave public fingerprints, so the real number is almost certainly much larger. But even that small signal is compounding.

The companies building real products for this layer are starting to post real numbers. The ones that haven’t — or that ran ahead of their receipts — just found out what the market thinks about the gap.

Zoom out.

Here’s the thing about this volatility: none of it changes the trade.

In February 2024 — before CRWV went public, DOCN found AI religion, and most people thought the word “neocloud” meant some sort of toilet paper — I laid out this thesis: the entire world’s computing infrastructure has to be rebuilt.

The reason: AI isn’t just a new application running on top of existing infrastructure. It’s a new architecture that replaces it. The traditional three-tier stack — frontend, logic, database — is collapsing into a single layer. The LLM is the logic and the database and (increasingly) the front-end. And that single layer runs on GPUs.

The bill to rebuild it is probably $2T+ and, while the hyperscalers will capture a lot of it, the gap is too big and expensive (and they’re going to be too slow and too general-purpose and they have their own $250B+ in legacy revenue to support) for the companies that need to run AI workloads at speed and at scale. That gap is exactly why the neocloud thesis exists. That gap is CRWV and NBIS, and it’s what DOCN is building toward and why AKAM just got a $1.8B vote of confidence.

This week’s earnings were a broader check-in on the thesis. The direction, for the companies with real proof, is not in question.

The trade, From the Porch.

The market is now moving on proof — so it’s going to be bumpy. Beat-and-punish quarters. Guidance-miss selloffs. RIF announcements that spook people who aren’t reading them right. Along with waking up to smart deals you didn’t expect that light a fire under your position. That is and was always the price of admission.

None of the volatility changed what these companies are building or what the demand curve looks like.

• DOCN and NBIS have the receipts. Hold them. Add on weakness.

• CRWV is the most frustrating stock in this group and probably the most important company. Three consecutive beat-and-punish quarters and the backlog is almost $100B. The capital structure is, well, anxiety-inducing, but the demand is not. If you have a long enough time horizon, the dip is a gift. (I’m buying.)

• NET is doing the right things and got punished for saying so too loudly. That usually corrects.

• AKAM is interesting now in a way it wasn’t last month. One contract isn’t a thesis, but it’s a first receipt. Watch.

• FSLY is the cautionary tale, not the opportunity. The stock ran on narrative. The narrative is unproven. Maybe they get there. But “maybe” isn’t the standard the market is applying anymore.

Every week that passes with more AI workloads running in production is another week where the infrastructure underneath them gets more essential and more defensible.

The market is trading quarters while the world is rebuilding compute. The stock charts will be bumpy, but the infrastructure layer for AI will be worth the ride.

Disclosed long positions: CRWV, NBIS, NVDA. Nothing here is investment advice. Do your own work.

Not (yet) a model company. Not (just) a cloud provider. (Sounds like a Britney Spears song. You’re humming it right now. Don’t lie.) Something aimed squarely at the act of running models in production, which is where the economics live. That framing matters more than the transaction itself, and it explains a pattern that’s been quietly compounding across the market.

What skipping this layer costs investors.

The popular narratives cluster(!) around models and APIs because they’re easy to see and easy to measure. OpenAI releases something, everyone writes about it. Anthropic announces a feature and entire stock universes get clobbered.

Meanwhile, there’s a whole layer of the stack that doesn’t announce itself loudly — and it’s where a growing number of real businesses are building.

On a call with Jim Fish from Piper Sandler and a dozen or so institutional investors last week, one of the recurring questions was: who exactly does DigitalOcean serve in an AI world? The confusion was genuine. (If I’m being honest, I struggled with it too.) The answer lives in this middle tier — which is exactly why it gets missed. And if you’re asking the question, you’re probably underweighting it.

The companies operating here mostly leave training to someone else and run models continuously, at scale, and in production. The difference between doing that well and doing it badly is increasingly the difference between a business with defensible margins and one quietly leaking money into a GPU somewhere.

That’s what skipping this layer costs: actual dollars. (Not a missed trend.)

Three tiers. One underappreciated middle.

Here’s how companies are actually using AI today.

• API users: teams building on OpenAI and Anthropic. This is what you hear about most. It’s the fastest path from zero to working. No infrastructure, no capacity planning, no late-night GPU therapy sessions. Totally valid until usage gets predictable and the per-call economics start to sting.

• Bare (or mostly bare) metal users: the model builders and large-scale operators (like, well, OpenAI and Anthropic) running training and fine-tuning on CoreWeave, AWS, and similar. Different problem set. Much bigger capital intensity.

• In between: a big, underappreciated chunk of the market — companies running their own models and forks of open models as a core part of their product and workflows. Customer support pipelines, document processing, internal copilots, agent systems calling other systems. The kind of thing that quietly becomes a real line item.

These teams are on platforms like DigitalOcean, Cloudflare, Together AI, Fireworks AI, Baseten, and Modal. Based on what we’re tracking across 7K+ companies in STAX (Porch Capital’s intelligence platform): OpenAI/GPT signals show up in 10% of stacks (766 companies); Anthropic/Claude in 5% (363). The detectable inference-layer — Modal, vLLM, Replicate, Baseten, Cerebras — clocks in at roughly 60 companies combined, with GPU Cloud (CoreWeave, Nebius, Crusoe) in fewer than 5. That 10:1 ratio almost certainly understates the middle. API calls leave public fingerprints. Self-hosted models and inference runtimes don’t — they’re inside the stack, invisible from outside. The 60 is a floor.

This isn’t a transient tier. It’s where a lot of teams find their groove. It’s also not one thing. It splits into two approaches that converge on the same customer from different starting points.

Same Destination. Different directions.

These platforms look similar from the outside. Dig in and differences show up.

DigitalOcean, for example, still feels like infrastructure first. GPU instances, layered tooling, Gradient (from its Paperspace acquisition) if you want guardrails. There’s a sense you’re operating an infrastructure system — a nicer one than before, but still an infrastructure system.

Together AI, Fireworks AI, and Baseten start from the model. Pick one, configure it, plug it in. The infrastructure exists but stays polite unless you go looking for it.

They’re circling the same customer from opposite directions. One path: machines first, add convenience and optimization later. The other: inference and tooling first, and then quietly build out the machines underneath.

Either way, the end state is the same. The interesting question is who gets there with better unit economics and more defensible positioning.

M&A is saying what product strategy is showing.

Lightning AI and Voltage Park merged in January, combining Lightning’s platform — built around PyTorch Lightning and used by 400K+ developers — with Voltage Park’s fleet of 35K+ H100, B200, and GB300 GPUs. The combined entity came out valued at over $2.5B with ARR exceeding $500M. The pitch: stop making developers choose between good software and cheap GPUs. Own both, bundle them, compete on the integrated margin.

CoreWeave completed its acquisition of Weights & Biases — reported at around $1.7B — doing the same thing from the infrastructure side. W&B brought training, evaluation, and monitoring tooling; CoreWeave brought (a fuck ton of) compute. The first joint product launch framed it as “metal-to-token observability” — infrastructure and application layer closing the gap.

Cloudflare has been moving in the same direction more quietly. Its acquisition of Replicate gives it a model execution and developer-facing layer on top of its global network — effectively pulling inference closer to the edge and bundling it with the underlying infrastructure.

Now Nebius is acquiring Eigen to add inference optimization on top of its compute layer.

Three deals. Three different starting points. Same destination: own the system that runs models in production end-to-end, not just a piece of it. The specific assets differ. The logic is identical.

Optimizing runtime performance.

Seen through that lens, the Nebius+Eigen deal makes more sense. (And just might explain the 14% Monday stock pop.) Nebius already had compute. Eigen brings inference optimization — software that determines how efficiently a model actually runs on that hardware. Combined, it’s an attempt to own both sides of the cost equation.

That matters because most of the margin destruction in this tier comes from inefficient inference — too many tokens, too much memory, too little batching. For the middle tier, the shift isn’t just about the integrated margin — it’s about the “Token-to-First-Byte” race. (It’s like Time-To-First-Token (TTFT), but different.) If you don’t own the optimization software (Eigen) and the metal (Nebius), you can’t hit the sub-20ms latency required for the next generation of voice-to-voice agents. If Nebius can control that surface, the economics of their platform get structurally better in ways that are hard to replicate by just buying more GPUs.

Worth noting: companies like Makora (disclosure: I am a strategic advisor / have equity) are building platforms that use LLM agents to auto-generate and inject optimized GPU kernels directly into PyTorch workloads. Pure inference efficiency, no hardware bundle required. And the creators of vLLM founded Inferact, with $150M of fresh-out-of-the-oven seed funding at an $800M valuation, co-led by a16z and Lightspeed.

Notably, Inferact raised before shipping an enterprise product. That’s a big bet that the inference runtime layer is a standalone, fundable, category-defining business. Which should tell you that the problem is real enough to support dedicated solutions, even before the infrastructure players try to own it through M&A.

The companies that win this layer won’t be pure infrastructure or pure tooling. They’ll be the ones that consolidate a full system, with enough vertical integration to build a cost advantage nobody else can easily close. The Nebius+Eigen deal is one version of that bet. (I’m long for a reason.)

Where this leaves the market, From the Porch.

A lot of the day-to-day AI work is happening in this middle layer. Products absorbing AI into their core workflows. Teams figuring out what it means to operate these systems over time. Platforms evolving — or merging — to meet them there.

The popular narratives will keep clustering around foundation models and API announcements. Those are easy to measure.

Watch the layer in between. It doesn’t announce itself loudly. It just keeps expanding.

Disclosures: I maintain positions in NBIS, CRWV, and NVDA. I am an equity compensated strategic advisor to Makora. I am an LP in an SPV that owns shares of Crusoe. I use AI liberally to help write posts, though the theses, believe it or not, are actually my own.