AI Semis Trade: From Arousal to Reality The past few weeks have been painful for AI semi investors. One of our mistakes early in the month was not realizing how much of an impact Anthropic’s ARR print would have on the AI trade, and not “sussing” out the set up heading in: lower distillation revs, OpenAI taking share in September, & META/MSFT off-ramps. These were all things that were reported on; if this was an earnings print, we wouldn’t have wanted to be long in the weeks heading into it. In hindsight — knowing what we know now, it was a perfect set up to reduce exposure into it. The upside of whiffing on that trade and the market’s renewed focus on Lab ARRs and financing? It has forced me (and other investors) to sharpen our pencils and better understand the different moving pieces of the trade, and how they play off each other. Let’s rewind a bit before diving into where we are now: For the first three years of the AI semi trade, and especially during the May/June go-go days, investors were in the throes of the honeymoon phase: maximum arousal, minimum scrutiny. Every piece of news was looked at with rose-colored-glasses. It’s just like the beginning of a new relationship when the infatuation is so strong that your partner can do no wrong. Even when she tells you she believes in an "open-source relationship," you somehow hand-wave it away and convince yourself it’s just going to soak up more of the frontier labs' excess demand. When your level-headed friends pull you aside to point out glaring red flags, like the fact that she’s running a shadowy circular-financing scheme with her roommates, you don’t listen; you just stop inviting them to parties. Before you know it, you’ve turned on Gavin Baker’s X notifications and forward SemiAnalysis to anyone who thinks you’re moving too fast, the ultimate wingmen who never stop hyping her up and telling you what you want to hear. “It’s way bigger than just coding, bro. She’s going to replace the entire global labor force.” Before long, the endorphins and dopamine completely take the wheel. Even if you carry the scars from past relationships — whether the early 2000s, GFC, 2019 SaaS, or the 2020–2022 Covid binge — the blinders go right back on. Those exes weren’t real supercycles and everyone knows there’s no dark fiber in GPUs, you tell yourself as you stare at GPU rental prices. You’ve even picked out baby names: Bel if it’s a girl; Claude if it’s a boy. But eventually the dopamine wears off, and you have to get to know the person you’ve been telling everyone is perfect. You begin to see the flaws and start asking practical questions about spending habits, expectations, whose money is actually paying the bills. Can we afford this lifestyle on those lab ARRs, or are we just annualizing the honeymoon and assuming the capital stays cheap forever? And is there enough power in ’28 to keep the spark alive? Things suddenly feel a lot more complicated, and we feel some investors longing for the simpler, more intoxicating days of May and June. But we see the end of the honeymoon as a chance for the relationship to mature. That means accepting where we are, adjusting expectations, and focusing our attention on where the strengths are rather than the flaws...

The Great Gross-to-Net Panic of 2026: OpenAI’s ARR in two charts & glass half empty vs. half full Glass half empty: People had thought $70B was net and the perceived dollar base just reset ~25–30% lower. OpenAI added ~$20B of ARR over August and September, not ~$30B. The first half was also weaker than people thought. The net run rate sat at ~$24–25B from February through June, and y/y growth slid to ~130% y/y by June, a bit decel from Dec 2025. That’s a lower starting point for YE26/YE27 net ARR targets & for valuation, and for the math on funding OAI’s compute commitments. Glass half full: The growth rate hasn’t changed. $30B → $50B net is the same ~70% in two months as $40B → $70B gross, and the monthly additions picked up into September. The y/y line bottoms at +131% in June, then climbs to +150% in July, +182% in August and +229% (~3.3x) in September. Had July really been $40B, September’s $50B would have meant y/y growth flat at ~230%. On the net base, it’s actually sharp re-acceleration of ~100% y/y in 4 months. Against Anthropic nothing changes, because the ~$70B gross figure is still the like-for-like comparison.

Not saying $FSLY can't continue to rally in this environment, but I ran the math and impact to #s# doesn't seem big at all. Agent browsing is heavy on requests and light on bytes. As FSLY's CEO put it this week, “agents aren’t watching videos yet.” Under our usage/pricing assumptions, each Muse daily user generates only ~1.5–5 cents in annual revenue before splitting traffic with Cloudflare. At 250M daily users, that’s ~$4–13M as sole carrier, or ~$2–6.5M with a 50/50 split. Stack every bullish assumption at 250M daily users and we get ~$35–60M, or 4–7% of FY27E revenue and roughly a 10% uplift to operating income. Reminder: FSLY fell 38% on May 7 after the last OpenClaw/agent run-up, when Network Services only grew a disappointing 11%. I know this different, jus pointing it out. Precedent: in 2021 Piper sized all of Apple’s iCloud Private Relay at $40–74M/yr, split across three CDNs. It never showed up as a visible step-change in FSLY's numbers. No position here, just know what you're getting into (Disclosure: I've been blown up on FSLY before)

Three years ago this week, we launched @TMTBreakout. Today we hit #5# on Substack’s Finance bestseller list. Thank you to everyone who’s been part of it. Three years in, and it still feels like we’re just getting started.

Notes from @OpenAI CFO Sarah Friar at GS Communacopia: Not a ton new on the #s#, but sounded bullish as expected... OpenAI CFO Sarah Friar used the session to put numbers on a business she said is now roughly half enterprise. The company entered 2026 at about 60/40 consumer to enterprise, set a goal of 50/50 by year-end, and reached it around mid-year because “the enterprise business has been on a tear”: in July, run-rate revenue grew about 20% month over month and enterprise grew 32% month over month “off a pretty big base.” The usage data behind that: a frontier firm, meaning a top-10% customer, now uses about 8x the tokens per user per week of an average customer, up from about 3x, and OpenAI’s own internal usage runs about 33x, which she offered as a preview of where existing customers are headed. Codex went from roughly 100,000 users at the start of the year to a last disclosed figure of about 25 million, and she said automated coding “is the thing,” with no developer left using AI merely to augment. Customer examples were concrete: Canva’s code is now “100% OpenAI” and its users create billions of images a week on OpenAI image models; Travelers took an AI claims-review assistant from eight states to nationwide in two months. Where customers ask next, in her walk-ins: cyber “very top of mind,” knowledge-work agents, and vertical transformation, with OpenAI investing in chip design, life sciences and financial services and hinting at a financial-services announcement in New York later this week. On the consumer side, with more than a billion weekly users, messages per user rise about 60% in the first six months and use cases double; Plus and Pro tiers are the year-to-date outperformers as users move “from asking to doing”; the ChatGPT and ChatGPT Work tabs merge by year-end; and the ads business hit a $1 billion run rate seven months after launch, “the fastest product to a billion-dollar run rate,” live in 40 countries with WPP and Dentsu on board and no AI-native ad format launched yet. Her framing of that opportunity: “if Google and Meta had a baby, you get ChatGPT,” high-intent search plus memory. The Astra launch anchored the product story. It was trained on 100,000 GPUs, “the largest training run we’ve ever done,” scores near 100% on ARC-AGI, on the math benchmark and on the Exploit Bench cyber benchmark, and Friar singled out computer use, agents working websites and applications on her behalf, as the capability that “blows my mind.” The economics she wants judged are cost per task, not cost per token: by an Artificial Analysis comparison, Astra needs 68% fewer output tokens than a rival frontier model to reach comparable results, and the 80% price cut on the smaller Luna model a few weeks ago produced a 10x lift in usage and the highest share on OpenRouter, above the next Chinese model. On open weights, she said they “have a place” but customers pay for the holistic stack, compute, data integration, firm context and enterprise reliability, and that “inference is not free”: Luna on Cloudflare is cheaper than GLM 5.3, so a frontier lab can serve inference below an open-source deployment. Pricing evolves subscription to consumption to outcome, and she would “love to get us away from token counting” toward sharing upside in verticals like chip design and life sciences, where GPT Rosalind is the life-sciences model and OpenAI’s own Jalapeño chip taped out in nine months, with the model running optimizations engineers did not have time to get to in the final 30 days before the design went to TSMC. On cyber, she called the idea that capable models can be kept “in a box” a fallacy: OpenAI ran Astra against its own environment to find and patch at machine speed, hosted about 300 CISOs in San Francisco last week, is giving trusted-access defenders more access with fewer guardrails, and sees “an incredible commercial opportunity” that carries a real governance burden. On compute, “we still feel really short,” with weekly trade-offs between training, research and serving, and the full-stack push framed as a cost loop: bigger training runs produce frontier models that train the cheaper child models, and low-latency serving for coders, voice, image and video all eats capacity. The stack is “a spectrum, not an either-or”: CSP partners, the in-house chip, a partner-built data center in Texas that OpenAI designed, and “more and more” self-build over time. Asked by the Goldman moderator how that ties to financing needs over three to five years, she said the hardest part of the job is projecting a business the world has never seen, sizing opportunities like cyber and the verticals, backing into compute needs and putting “shovels in the ground” two to three years ahead; the past two years’ criticism for over-investing “really paid off this year,” and the discipline is ROI: for each model family they look back at revenue to date plus 12 months forward against the compute invested and ask whether it is “not just cracking positive, but like hell positive.” Revenue per gigawatt rising and the cost curve falling are what expand gross margin, in her telling. On applied research versus visible demand, mission comes first, with investment across voice, video, images and world-simulation models, though she noted nobody could have modeled Codex’s scale a year ago. Her priority list for the next few years, in order: frontier intelligence; staying on the Pareto frontier so intelligence is “mass available”; driving compute cost down so revenue per GPU rises; the consumer base as a flywheel into enterprise; and enterprise capabilities, with cyber the focus of the next 12 months and horizontal office-of-the-CFO, CRO and CMO products alongside vertical specialization.

Good chart from @TMTBreakout July accelerated both MoM and YoY for the sum of OpenAI and Anthropic. Some chance Anthropic shifted from gross to net reporting for ARR. Meaning the 65b is a more conservative metric than the $47. Smart thing to do before an IPO. Openrouter data shows that AI broadly has accelerated further over the last 3 weeks. Open source taking share is positive for AI infrastructure demand.

Chinese optical stocks are plunging on rumors that the U.S. will impose sanctions, while U.S. optical stocks are plunging on fears that it won’t. What an irony. This is from the @TMTBreakout newsletter.

Some thoughts on recent price action in Tech: Post-earnings action continues to be a lot better in internet & software. Investors there are still showing an appetite for buying names up post-earnings, especially where numbers accelerated and the narrative improved, even if just marginally. $PLTR +2%; $TEAM +2%; $TWLO +3.5%; $U +25bps; $UBER +4%; $SAP +1%; $MSFT +1%; $ABNB +4%; $TTWO +2% are a few of the names that printed solid earnings and have follow-through. This shift in flows out of semis started with $s flowing into hyperscalers a couple weeks ago and has now expanded. We even started to see some willingness to buy the dip on misses, with $DDOG +11% and $FIG +9% today, while something like $AKAM +6.5% didn’t stay down long. Compare that with the anemic follow-through in AI Semis: $ANET following its big beat, $STX, or $ALAB selling off despite the big Sept Q guide. The weaker prints like $INTC and $SNDK also continue to struggle to find a bid. There’s admittedly some cherry-picking in the attached table showing how stocks have reacted following prints, but in our defense, there are a lot of cherries. To put it succinctly: with the AI semi vibes & narrative remaining choppy and some of the more favored parts of the trade (memory, CPUs, storage, etc.) seemingly rangebound for the time being, investors appear increasingly comfortable putting $ to work in idiosyncratic ideas outside the space in names where the earnings narratives are moving in the right direction, numbers are beating, valuations have come down, names have underperformed YTD, and sentiment is tilted to the left. We thought some of those flows would also go to $NVDA, but even that is lagging. Today felt like that risk appetite broadened another notch outside AI semis, with investors willing to spread dollars further down the bench, buying misses and revisiting high-quality underperformers without an obvious near-term catalyst, as the $NFLX/ $SPOT moves illustrated.

For those thirsty for a positive AI semi px action tea leaf, $TSM -- which typically likes to do a sell the news on earnings -- actually finished higher than it opened, which is different than $AEHR, $ASML, & Samsung EPS T + 1.

A richness of great insight on the @BG2Pod from @GavinSBaker ...too many to choose from but here's a few good ones: “Something very important on open source: there’s this belief that it’s bearish for AI. It’s actually — maybe bearish for the frontier models — but really bullish for compute and hardware. If the frontier models are capturing less of the margin, then you’re going to spend more on compute. The better open source does, the better it is for compute providers..." “Two things can be true. The majority of economic value may continue to accrue to the frontier — and man, has it ever accrued to the frontier thus far. And the majority of tokens consumed in the world may be open source — and they are today. I think this current state is likely to persist.” "There is a belief that these data centers are commodities. I do not share that belief. In the same way that Elon was able to re-engineer a rocket from first principles and make it reusable, he engineered an electric car from first principles — I think he looked at data center design from first principles and designed something fundamentally different." "My understanding is that Cursor and Anthropic have more tokens of proprietary coding data than anyone else — and each have more tokens of proprietary coding data than exist on the public internet. Cursor used Kimi K2.5, used their own private data, did some RL, some supervised fine-tuning, and got a really good model. Then they spent three weeks in the Colossus 2 cluster and got a model that 12 days ago was Pareto dominant with Composer 2.5. It suggests the Cursor data is very valuable for coding, and that XAI/SpaceX has a shot at being a real player in coding.” “Nobody has run Mythos for a year continuously, and we may never know how smart each generation of models actually is or was — because we don’t have time to appropriately evaluate their intelligence before the next model comes out. This is a profound statement...Imagine Albert Einstein had just thought about fundamental physics 24 hours a day. He doesn’t have to eat, doesn’t have to sleep, never gets old, never has diminution of intelligence — and he thought for one year. We might already have solved a lot of these intractable problems. My takeaway was: however bullish I was on compute before, I’m just a lot more bullish.” On $NVDA: If all of his customers are going to compete with him, then why not compete with his customers? He has his own models that are really, really good — Nemotron 3.1 was really cool from a compute-efficiency perspective, and he’s always careful to release small models so as to not tread on Anthropic, OpenAI, Google’s toes. But that is a choice he is making. If the economics change, I think Nvidia can join the frontier and become one of the world’s largest cloud computing companies much faster than people think.” “Clark’s analysis shows that XAI’s deal with Google for cloud computing generates more operating profit per gigawatt than Anthropic, than Meta, than Google, than OpenAI. Freda calculated a 55% IRR on Colossus 1. If you can borrow money at 6, 7, 8% and invest in something with a 55% IRR — I’m not the most sophisticated thinker, but that math maths.” “I always imagine stocks as runners. In ‘22 that runner had gone downhill — it had a lot of energy. The market, particularly in the last two months, has run up a very steep hill. A lot of these stocks — forget climbing a mountain — they’ve gone straight up a cliff. They’re tired. They need to rest. Do they just rest at the top of the cliff, hang out in their harness, or do they need to go downhill for a bit? We’ll see. I do see a lot on X about finding the next bottleneck — I think that was the last game. That game is over.”

@TMTBreakout has built an incredible community that is starting to realise its full potential on Slack. @FundaAI just wrapped an AMA and you can expect Apptopia on Wednesday, April 22nd, 11am ET and Doug O’Laughlin on Friday, April 24th, 10am ET.

Cool little milestone for @TMTBreakout: just became a top 10 finance Substack for the first time ever. Thanks to the great writers on this list like @citrini @agnostoxxx @jbulltard1 who have helped the Substack platform grow and become what it is