Expo Research (@ExpoResearch) header
Expo Research (@ExpoResearch) profile picture

Investment strategist. Views are NOT investment advice.

London, England
Joined July 2019
Photo by Expo Research (@ExpoResearch)Photo by Expo Research (@ExpoResearch)Photo by Expo Research (@ExpoResearch)Photo by Expo Research (@ExpoResearch)Photo by Expo Research (@ExpoResearch)Photo by Expo Research (@ExpoResearch)Photo by Expo Research (@ExpoResearch)Photo by Expo Research (@ExpoResearch)
Expo Research retweeted
SemiAnalysis has HBM pricing in 2027 falling double digits and holding flatish afterwards through 2030. They have Nvidia shipments (their previous estimates not relative to today) getting nuked from space. 2027 falls double digits. While the out year falls above 50%. (Aren't blaming volume cuts on Jensen leaving) AMD faces a similar issue with double digit drops in 2028. Out years have them falling above 50%. (their previous estimates not relative to today) ASIC programs face barely any volume cuts. Due to AI labs preference for superior TCO. But they target Mediatek due to weaker financing (I quibble with this given Google will buy enough internally and Anthropic can avoid em). Overall SemiAnalysis take is total accelerator revenue only grows 50% from 2027 to 2030. Primarily due to memory price cuts. Total logic silicon shipped still 3xs from 2027. (You should read this as logic Foundry is a much better business than memory LOL). Anyway my take from this is SemiAnalysis finally puts away the pie in the sky view. If you think this is true then idk why we need TeraFab and spot pricing is gonna be bad. Labs need to accelerate revenue or we are in a bad spot. The impact to Intel is be careful on memory. Also getting ASIC customers is more important than ever before. GPUs are cooked as a business selling to labs. $INTC $NVDA $AMD
〈下調加速器需求與出貨量展望;NVIDIA 承受重壓〉 1、下修幅度
1)SemiAnalysis 在機構版內容中下調了未來 AI 晶片出貨量、相關支出和用電需求的預測,2028年起的調整最明顯,NVIDIA( $NVDA)是受影響最大的供應商。
2)2028年全球 AI 加速器資本支出預測從約1.05兆美元降到7,580億美元,下調28%。
3)2030年預測從約1.53兆美元降到9,230億美元,下調40%。
4)報告截圖說明,下修原因包括出貨量下降、價格下降,以及調整了2029和2030年未來加速器的單晶片功耗假設。 2、下修的三大原因
1)重新設定生態系的總 Token 收入預期:Frontier Labs 第三季淨新增 ARR(年度經常性收入)成長放緩,所以下調新增供給的預估,讓收入能持續覆蓋整體算力成本。SemiAnalysis 說自己對實驗室收入的預期仍比市場共識樂觀,也看到很多值得部署前沿智慧、投資報酬率高的領域,但資本市場參與者的平均預期對基礎建設融資非常關鍵。
2)資本市場收緊:債權和股權投資人都不願投資 AI 建置相關的資本支出,沒有超大規模雲端業者承租的案子尤其如此。這會減少資料中心用地、電力和廠殼開發,可供部署系統的電力供給因此下降。
3)美國期中選舉結果:原本規劃在2027年下半年到2028年上半年建置的5.8GW 美國資料中心供給被延後,監管延遲也在減少2028年的資料中心總容量。SemiAnalysis 說這確實讓近期算力設備出貨量下降,但前兩項因素更重要。 3、衝擊不平均:NVIDIA 首當其衝
1)SemiAnalysis 的架構是,客戶會優先選擇效能、TCO(總擁有成本)和每瓦效能最好的晶片,再疊加融資支援;在這些指標上排名最差的平台和晶片,會被不成比例地削減。
2)結果是 NVIDIA 受影響最大,承擔了大部分的支出削減。
3)ASIC 方面,聯發科(MediaTek, 2454.TW)的 TPU 項目也受影響,因為融資能力比 Broadcom( $AVGO)弱。
21
15
2
118
24,200
Expo Research retweeted
As someone who’s hoped that compute will be a bottleneck to RSI, the intuition in this footnote hit me like a sledgehammer
The avalanche that hit math is coming for AI development. dwarkesh.com/p/rsi
71
116
37
2,828
461,050
Expo Research retweeted
I went all-in on $PEPE on Feb 24, 2024. Two days later it broke out and ran 10x+. I've watched this chart every day since, and I think it's doing the exact same thing again. 🧵
16
38
2
312
22,305
Expo Research retweeted
$ANT guided a 35%+ long-term operating margin. I think the near-term debate will be about compute costs. ANT's official long-term guide is an adjusted operating margin of 35%+, with compute and infrastructure (C&I, the chips and data centers that run its models) at ~45-50% of revenue. That implies gross margin of ~50-55%, against about 42% in Q3 2026. I find that interesting. It assumes the compute shortage eases only partly, or that once compute is abundant ANT will not sell its models at huge margins. I also wonder whether sales and marketing, and personnel and other operating costs, really stay at ~7-10% of revenue each. Longer term, if revenue scales massively, the two combined could fall to a low single-digit % of revenue, because these costs should be very scalable. Before then, I think the market will argue hard over C&I as a cost of revenue. The trajectory is against ANT: C&I rose from 48% of revenue in Q2 to 58% in Q3, and could get worse near term. The incoming Democrat-led Senate and House will add restrictions, and states and cities, including Texas under its new Democratic governor, are opposing new compute buildout. So the long-term picture may be great, but near term the market may focus on compute costs and a worsening gross margin. If ANT defends gross margin, growth slows. If it maximizes growth and customer value, gross margin comes down. Either way, the near-term trajectory looks worrying.
5
5
1
55
6,852
Expo Research retweeted
near instant validation for my prediction teapawt.net/yrechtman/status/21187…
We raised our $750 million Series D at a $20 billion valuation from @benchmark, @generalcatalyst, @nvidia, @salesforce, and @zoom. @mercor scaled from $1M to $3B in run-rate revenue in 27 months, the fastest in history to $3B. For 2 years, we've organized human intelligence to build evals and training data for every frontier lab. The labs are the leading indicator for AI startups, then the Fortune 500. Our growth has accelerated over the past 6 months as enterprises start evaluating and training their own models. For many, owning their intelligence is existential.
8
5
1
253
29,033
Expo Research retweeted
We raised our $750 million Series D at a $20 billion valuation from @benchmark, @generalcatalyst, @nvidia, @salesforce, and @zoom. @mercor scaled from $1M to $3B in run-rate revenue in 27 months, the fastest in history to $3B. For 2 years, we've organized human intelligence to build evals and training data for every frontier lab. The labs are the leading indicator for AI startups, then the Fortune 500. Our growth has accelerated over the past 6 months as enterprises start evaluating and training their own models. For many, owning their intelligence is existential.
302
261
157
3,582
1,351,653
Expo Research retweeted
Compute efficiency
Broad-based adoption metrics across customers
1
5
1
75
25,188
Expo Research retweeted
Anthropic published its IPO roadshow slides and presentation today. Below, I provide some highlights. Scroll through, and I've provided a link to all the slides at the end. 👇 1. $ANT is growing rapidly without burning cash and already has a fortress balance sheet.
10
26
1
160
32,432
I see the new criticism of Anthropic's S-1/A is that ANT "inflated October ARR by slashing token prices, which will result in compute margins falling off a cliff in Q4" as "ANT is now selling dollars for 50 cents!"... Well, no, lol. The idea that ANT's compute margins declined significantly MoM in October insinuates that ANT was able to bring online multiple gigawatts of compute immediately on October 1st.. which obviously didn't happen. Their own disclosures combined with some other (conservative) assumptions suggest compute margins will be flat to up in Q4-26. $81 billion of October ARR implies ~$20 billion of Q4-26 revenue assuming zero NNARR in November and December (which seems like a pretty bearish assumption given October NNARR was $10 billion!). We can estimate Q4-26 compute and infrastructure spend as: - $5.2 billion of disclosed contractual obligations due 2026 as of Sep. 2026 (e.g. to be incurred in Q4-26 with Amazon, Google, etc.), plus: - $7.2 billion estimated SpaceX spend (see below) ... for a total $12.3 billion of Q4-26 compute and infrastructure expense. So assuming zero NNARR in November and December.. Q4-26 revenue of ~$20 billion less ~$12 billion of compute and infrastructure spend implies ~$8 billion of compute profit dollars and a 39% compute margin.. which is relatively unchanged from Q3-26, and still pretty healthy.. Obviously in a more realistic scenario in which November and December ARR continue to grow, Q4-26 compute margins would likely exceed those of Q3-26. Note: $SPCX disclosed Anthropic (Customer B) as their largest customer in Q3-26 at 34.5% of total revenue ($14.6 billion).. so $5.0 billion of SpaceX's Q3-26 revenue came from ANT. We also know Anthropic's cloud deal with SpaceX ramps to $3.1 billion per month in December 2026, which implies $9.3 billion quarterly spend. The analysis above uses the average of Q3 2026 spend ($5.0 billion) and the full ramp ($9.3 billion), or $7.2 billion, as a proxy for ANT's Q4-26 spend with SpaceX. $ANT
2
12
2
102
19,437
Expo Research retweeted
“Our Run-Rate Revenue was approximately $81 billion as of October 31, 2026, up $10 billion from September 30, 2026.” $10 Billion of incremental revenue. In a single month. Which included a 40% price cut. Uncharted territory.
104
81
34
1,508
211,596
Expo Research retweeted
🆕 Anthropic $ANT IPO Roadshow (57 slides) available on RetailRoadshow: netroadshow.com/custom/IPO/A…
6
79
4
520
111,077
Expo Research retweeted
A bit of Anthropic roadshow color, according to people familiar with the matter - Expected to price next Wednesday, begin trading on Thursday (Nasdaq-listed) - I’m told they’re already covered from institutional investors at the $70-$78/SH range on an allocable basis (meaning, extracting over-inflated orders). So, they’re already technically oversubscribed - They’re estimating that about 20% of the offering will go to retail - I’m told the institutional demand is even stronger than SpaceX at this point in the process (although SpaceX was raising more from the public markets because they didn’t have the cornerstone component) $ANT
4
15
7
86
21,947
Expo Research retweeted
🚨 ANTHROPIC $81B ARR AS OF OCTOBER 31, 2026 This is a $10B (14.1%) increase from September!
17
17
5
381
67,427
Expo Research retweeted
ANTHROPIC: GOOGLE, NVIDIA, AMD, BROADCOM, MICRON, MICROSOFT, TSMC TO INVEST $26B THROUGH PRIVATE PLACEMENT AT IPO PRICE ANTHROPIC: CORNERSTONE INVESTORS WILL SELL 23,333,331 SHARES IN THE IPO
4
9
4
176
136,204
Expo Research retweeted
Claude Fable 5.2 sets a new record on the Epoch Capabilities Index (ECI) at 174. The previous high was GPT-6 Astra’s 169. Fable 5.2’s score is above what the pre-Astra trend for reasoning models would have predicted. This is notable given recent discussion of AI contributions to AI R&D, and we hope to release more work in this area soon. This is based on testing we did prior to launch using only our own benchmarks. The ECI is continually updated as more benchmark results become available, and may change in the future.
70
186
47
1,607
179,411
Expo Research retweeted
I've summarized where things stand at the AI frontier: - Unreleased models (Anthropic's Models 3 and 4) are now half-way between AI2027's Agents-1 and 2 (graphs fixed for the bug flagged by the AI2027 authors). @DKokotajlo please let me know if I missed something. - We're right on track with AI2027's prediction of an automated coder (AC) for April 2027, wild. Public models are now massively lagging unreleased models (1 day vs. 1 week of autonomous high-quality work). Fuelled by RSI, the gap is huge and growing. - This trajectory, if continued, means a world-changing intelligence explosion in 2027, so the public needs to know. I highly recommend reading more on the topic to grasp what's coming.
83
306
76
2,445
425,739
Expo Research retweeted
Yes, you need exponential increases in spending for linear increases in ECI. But linear increases in ECI translate to exponential increases in productivity at RnD, potentially allowing an intelligence explosion Exp increase in spend -> linear increase in ECI --> exp increase in output at AI RnD
Updated version!
15
26
6
305
22,108
Expo Research retweeted
Updated version!
In 2025, AECI increased by 4.05 points per $ bn of R&D. In 2026, this was 1.49 AECI points per $ bn of R&D. So progress per R&D dollar fell by 63%, even while progress per year doubled.
42
101
57
846
305,313
AMD's expected to quadruple in 3y. $AMD
9
15
1
221
13,861
$INTC Intel is already deep in the memory integration game and the $TSM 3D DRAM rumor highlights why both are now critical. While TSMC is reportedly exploring actual 3D DRAM production at Nvidia’s request amid the shortages, Intel has been building the other side of the same problem for years. Intel isn’t jumping back into commodity DRAM manufacturing but it is attacking the memory wall through advanced packaging and new architectures: - Strong focus on stacking memory directly with compute (Foveros, EMIB, hybrid bonding). - New concepts like XBM (an interposer-less stacked memory approach) and ZAM to improve bandwidth, power, and cost versus traditional HBM. - Hiring former SK Hynix leadership into senior Foundry and packaging roles. - Treating advanced packaging as a major growth engine, not just a supporting technology. This is the inverted version of TSMC’s potential move. TSMC dominates pure logic foundry and is being pulled deeper into memory. Intel is an IDM leaning harder into foundry services while using packaging and system-level design to solve the same AI bottlenecks. If TSMC can execute on 3D DRAM without diluting its logic leadership, and Intel can deliver competitive packaging and process nodes (18A/14A) at scale, both become even more indispensable to the AI buildout. The risk for each is execution across a broader footprint than they’ve traditionally optimized for. Nvidia (and the hyperscalers) are effectively forcing the entire supply chain to blur the old lines between logic, memory, and packaging. Whoever solves the memory wall best, whether through pure process leadership or system integration, wins outsized leverage.
TSMC WEIGHS MAKING 3D DRAM AT NVIDIA'S REQUEST $TSM has reportedly begun drawing up plans to manufacture advanced 3D DRAM chips as AI-driven memory shortages push chip designers to seek additional supply, per The Information. Unlike conventional DRAM, 3D DRAM stacks memory cells vertically to increase density. The move could put TSMC in direct competition with Samsung, SK Hynix and Micron. No production timeline has been set. TSMC is considering building new fabs or converting an existing facility, but says its focus remains on its core logic foundry business.
3
2
76
7,359