Weekly Recap: AI Stocks Fall, Texas Turns Blue, and Google Bets on Bloom
Deleveraged market, crowded shorts and reopening buybacks set up squeeze into year end
View original →Digestible, Institutional Grade Investment Research.
Cite asCryptoQuant has tracked the directional calls on NVDA by @asymmetricalbets on Substack since Feb 2026: not yet scored on NVDA (as of Nov 21, 2026).
2 NVDA calls on record, a thin history.
Relative view by default: fewer than 30 NVDA calls or days on record.
Sentiment timeline shows relative sentiment within this analyst's history. A perma-bull showing 8 bullish : 2 bearish in a bear market sets that as their baseline. If you notice any errors, claim @substack:asymmetricalbets to submit corrections.
Hover the chart to read the thesis of that period · click a day to pin it
Weekly Recap: AI Stocks Fall, Texas Turns Blue, and Google Bets on Bloom
Deleveraged market, crowded shorts and reopening buybacks set up squeeze into year end
View original →Anthropic’s First Filing
Anthropic filed its S-1 this morning under the ticker ANT. The company says it’s building “a country of geniuses in a data center” and sits at the center of the AI investment cycle. Its demand for compute helps shape spending across chips, cloud infrastructure and data centers, and headlines about its models move semis stocks within seconds. Whether that status holds or not, a company pulling that kind of gravity deserves a close read when it opens its books for the first time. We read through the filing to answer three questions: How strong is the underlying business? What will its compute commitments cost? What will public shareholders actually own? Before we begin, this Substack is for informational and entertainment purposes only and does not constitute financial advice. Past performance is not indicative of future results. Never invest more than you can afford to lose. Our writers may hold positions in the securities discussed and may buy or sell them at any time without notice. Table of Contents: Anthropic’s Thesis The Financials Infrastructure Cost The Profitability Question Unusual Ownership Structure What To Watch 1. Anthropic’s Thesis The S-1 opens with: “For the first time in history, intelligence is something humans can build” and Anthropic points to more than $40T in global knowledge-work spending as its potential market. Anthropic describes itself as a research lab first, but the business model that funds it is based on usage. Anthropic earns revenue through consumption charges and subscriptions, with usage measured in tokens and priced by which Claude model customers pick. That consumption pricing made up 80% of H1 2026 revenue. Claude is distributed through AWS, GCP and Azure, with Anthropic booking marketplace revenue gross and paying each cloud provider a distribution fee. Anthropic sells two main apps on top of its models. Claude Code writes software, and Claude Cowork handles office work like research and drafting. Claude Code shows how a more intensive AI product can increase customer spending. In Q3, Claude Enterprise users who didn’t use Claude Code spent under $75 a month on average, while the top 5% of Enterprise users who did spent about 20x that. Moreover, Anthropic also runs Claude on itself. Over 90% of the code added to its own codebase in September 2026 was written by Claude. The company credits that in part for keeping the team lean, with about 5,300 employees bringing in over $6.6M of revenue each over the last twelve months. The company expects what it calls “Powerful AI” to arrive within a few years, capable of working autonomously for hours or days at a time, and believes demand for that kind of intelligence will outstrip supply for the foreseeable future, constrained only by compute availability (as opposed to what Elon Musk recently said on X). 2. The Financials Revenue went from $41M in 2023 to $4.6B in 2025, and then to $16.3B in the first half of 2026 alone. Q3 came in around $16.9B, based on the filing’s disclosure that September quarter revenue was 86x bigger than the Q4 2024 revenue. That’s roughly 47% growth over Q2, a slowdown from the 144% jump the quarter before. Revenue grew faster than the two largest expense categories.. Compute and infrastructure ate 251% of revenue in Q1 2025 and by Q2 2026 that ratio was 48%. Personnel dropped from 41% to 16% over the same period. Q2 2026 was the only quarter of the six disclosed to generate a GAAP operating profit. It posted $974M in operating income on an 8% margin. On an adjusted basis, which strips out stock comp and the Bartz settlement (a $1.5B copyright lawsuit booked in Q3 2025), that margin was 19%. And the most important part of the filing, which is the widely cited $80.7B net loss in H1 2026 looks alarming, until you open it up. $72.4B of that is a non-cash accounting charge on convertible notes and SAFEs (simple agreements that convert into equity at a future event like an IPO), driven upward on paper as the company’s valuation rose. Another $8.8B came from revaluing note purchase agreements with Google and Amazon. So, the actual operating loss for the half was $693M, and operating cash flow was positive at $3.9B. 3. Infrastructure Cost Anthropic manages its compute as a shared pool that serves both its own research and its paying customers. The infrastructure spans Trainium, TPUs and NVIDIA GPUs today, with AMD capacity expected in 2027. Our reconstruction of the footnotes puts disclosed commitments at roughly $853B: → $265.6B of purchase commitments at June 30 → $321.4B of leases not yet commenced and approx. $266B of subsequent commitments The filing’s consolidated commitment table is still blank, so this is our reconstruction rather than a total presented by the company The largest single counterparty is Broadcom at roughly $161B in equipment lease rent, followed by SpaceXAI at up to $120.5B in NVIDIA capacity, Google at $114.5B in TPUs, and Amazon at $110B in Trainium. These obligations run as far out as 2045, and the filing says they “generally must be satisfied regardless of our revenue or utilization,” so if demand for Claude slows or a chip generation underperforms, Anthropic still owes the payments. The terms differ by deal though: the SpaceXAI agreements allow cancellation on 90 days’ notice after initial non-cancelable periods ending in early 2027. The financing behind these deals works the same way we’ve seen with NVIDIA and the neoclouds. The companies selling Anthropic compute are also helping pay for it. Broadcom, Google, Amazon and AMD each play at least two roles at once: supplier, lender or shareholder, and sometimes all three. has a convertible note facility of up to $42B used to finance the rent on its own equipment. Interest is added to the debt balance rather than paid in cash (PIK) at a standard rate of 9%, rising to 13% when certain revenue conditions aren’t met. Those notes can also convert into shares, which would dilute public shareholders. has a $20B note facility tied to revenue targets and Amazon has a $15B facility that opens in $5B steps as compute milestones are met, on top of $6.2B of older notes it already holds Some of these contracts are linked, because if Anthropic misses a payment on one deal, other lenders can demand their money back at the same time, which could cut off access to compute. 4. The Profitability Question Anthropic posted its only profitable quarter in Q2 2026, but the filing is upfront about what comes next. New compute capacity arrives in large steps, and the expense hits before the revenue that uses it. H1 2026 gross margin was about 85%. That measure excludes model development spending and cloud distribution fees though, both of which matter for the full picture. Blow gross profit, there are larger expenses: → Total R&D, including research compute and personnel ($9.74B, or about 60% of H1 2026 revenue). Models released in the six months through September drove more than 80% of Q3 revenue, which means continued research spending isn’t optional if the company wants to hold its revenue base. → Channel fees paid to AWS, GCP and Azure (which drove sales and marketing to 22% of revenue in Q1 2026 before dropping to 18% in Q2) → Stock-based compensation that will spike after the IPO If the IPO had closed on June 30, q2026, $2.9B in stock comp would have hit the books immediately. Another $23.7B sat unrecognized at that date and is expected to flow through over roughly 2.1 years. The filing gives no profitability target and no timeline for sustained positive margins. However, the S-1 says that the company’s ability to keep raising capital “depends in part on demonstrating a credible trajectory toward profitability.” Operating cash flow was positive at $3.9B in H1 2026, and Anthropic sat on $131.7B of cash and investments as of June 30. Anthropic also earned $762M of interest income in Q2, recorded separately from its $974M operating profit. 5. Unusual Ownership Structure Anthropic has five classes of stock, and the filing itself says this structure has “no direct public company precedent.” Under the planned IPO structure, the founders’ entity will hold a fixed 50.1% of voting power through a single Class F share controlled by their Founder LLC. Public shareholders get Class A, which carries one vote per share but is collectively capped at 49.9% of the total vote. Class N shares, held mostly by Amazon, Google and the Public Benefit Fund, carry no votes at all. Board control sits with the “Long-Term Benefit Trust,” which elects a majority of directors through one Class T share. The trust has five authorized seats with four currently filled, and its trustees include Ben Bernanke, the former chair of the Federal Reserve who led the U.S. central bank through the 2008 financial crisis. The filing states what this means for public investors: Class A holders “generally will not be able to affect the outcome of any matter.” The lock-up expires in stages rather than on one date. Employees (aside from the founders) can sell 10% on day one, with most of the rest unlocking after full year and Q1 2027 earnings. The seven co-founders wait the longest, with 29% freeing at Q1 2027 results or day 180, and the remainder at Q3 2027 results or day 365. The filing leaves share counts per group blank, so we can’t size the actual supply hitting the market at each date yet. 6. What To Watch Read more
View original →Weekly Recap: AI Stocks Crash Further, SpaceX Guides 10 GW, And What We Bought And Sold
Last week we walked through the AI sell off day by day and argued that the macro outlook changed with the inflation prints, but at the stock level the damage was mostly sentiment driven. CoreWeave was the one exception where credit stress looked real. Since then, it got worse. Goldman Sachs’ prime brokerage team told clients that October 22 was the first time since the COVID crash that all three major hedge fund strategies they track lost more than 1% in a single day. Long/short, systematic, and multi-strategy funds all got hit at once. Hedge funds focused on tech are down roughly 10% this month and funds invested in Asia are down about 9%. Reuters reported that the selling triggered institutional deleveraging events across global markets. Meanwhile, retail was the one that kept buying. According to separate interviews from Citadel and JP Morgan, retail clients have been buying the dip in semiconductors, neoclouds, and memory stocks at the largest rate in 16 weeks. Source: Reuters Earnings season started on top of all of this. SpaceX beat across the board, went up 10% after hours, and gave it back within 48 hours. Meta beat on every metric, gapped up 8%, and opened flat. Every surge is getting eaten and even though companies are delivering, the market won’t pay for it. The macro backdrop is doing most of the work: → Oil swung about 6% in four days on Middle East escalation and Brent is sitting near $112 → The FOMC meets Oct 27 to 28 and announces its decision Wednesday. The blackout runs through Oct 29, so there’s no Fed commentary until the decision. December hike odds are at 73%, and it seems to be the only appropriate decision. As long as that’s the case, strong earnings aren’t enough to hold a bid. In order to find out what to do in this environment, we went through everything that reported and everything that moved this week to figure out what the numbers actually say underneath the so far weak price action. Before we begin, this Substack is for informational and entertainment purposes only and does not constitute financial advice. Past performance is not indicative of future results. Never invest more than you can afford to lose. Our writers may hold positions in the securities discussed and may buy or sell them at any time without notice. Table of Contents: Saudi Arabia and the Oil Swing SpaceX Q3 Earnings Super Micro Prelim Results NVIDIA DSX Pod (GTC Berlin) Meta Q3 and Muse vs Grok Bot Sony and the GTA 6 Trade Anthropic S-1 Bloom Energy Q3 Credo and the MaxLinear AEC Threat What To Watch [1] Saudi Arabia and the Oil Swing The week actually started with some positive news, Qatar and Pakistan floated a deal on Tuesday: Iran gets the interest on $6B of its funds frozen in Qatari banks, and the IAEA gets confidential access to Iranian nuclear sites. Oil dropped 3% intraday on the report and Brent touched $106. The deal fell apart the same day. Iran’s parliament speaker told Pakistan’s army chief that Tehran had nothing new to offer and neither government confirmed the proposal. The only source was Al-Araby Al-Jadeed citing unnamed officials. On Wednesday, a drone aimed at Aramco’s Shaybah oilfield was intercepted, killing two. Hours later a strike on King Fahd airport in Dammam killed two more and injured eight. The State Department ordered staff and families out of Saudi Arabia and then Thursday brought Houthi claims on Abha airport, oil sites in Yanbu, and the Saudi defence ministry. The IRGC said it set four tankers on fire in the Gulf. King Fahd airport after the attack One unconfirmed peace headline moved oil 3% down, and two days of strikes pushed it more than 6% back up. Brent went from $106 on Tuesday to over $112 by Thursday morning. A lot happened in the span of three days, and the market reflected that, KOSPI fell 3.52%, Nikkei dropped 2.65%, and S&P futures were off 1%. The 10-year yield fell about 7bp to 5.075% as investors bought bonds, but oil near $112 keeps a December hike on the table. [2] SpaceX Q3 Earnings SpaceX reported Tuesday after close. Revenue came in at $14.6B, up 161% from a year ago. Net income was $4.4B against a $1.1B loss in Q3 2025. The beat over consensus was about $1.6B. The AI segment did the heavy lifting because “Super Intelligence” revenue hit $7.56B, up about ninefold and now more than half of total revenue. That includes $608M from Cursor, which SpaceX acquired for $60B and closed during the quarter. Moreover, Nameplate capacity sat at 1.7GW at quarter end, with nearly 2.2GW online by end of October and a target of 2.5 to 3GW by year end. Q3 capex was $19.6B and Q4 is guided to $35B to $40B. The CFO said current rental prices let SpaceX recover the cost of new compute in under a year. Both SpaceX and Credo (of which SpaceX is the biggest customer) went up around 10% after hours, gave it all back within 48 hours, and was down 3% early Thursday with the broader selloff. BofA raised its target to $275 and Cantor Fitzgerald moved their PT higher to $345. [3] Super Micro Prelim Results Super Micro posted preliminary Q1 numbers on the same day as SpaceX. Revenue guidance jumped to $18.5B to $19.5B, up from a prior range of $14.5B to $15.5B. Bloomberg consensus was around $15B so even the low end of the new range sits 19% above the top of the old one. Gross margin was the bigger surprise because the company guided 10.4 to 10.8% last quarter and estimates from Reuters had it at 10.15%. The prelim came in at 12.0 to 12.5% and management credited customer and product mix for the improvement. The stock jumped 12 to 13% on Wednesday and pulled the server names with it. Dell rose about 6%, and HPE 5%. Samsung and SK Hynix gained 4 to 5% in Seoul on the memory read through. Rosenblatt kept Buy and raised its target to $57 from $51. However, expectations were low coming in. SMCI had said last quarter that fiscal Q4 would land near the low end of guidance. The numbers are preliminary and unaudited, and the company has a history of volatile results, so you should take this information with a grain of salt. Full results and the call come Nov 3. [4] NVIDIA DSX Pod (GTC Berlin) On Wednesday, at GTC Berlin, NVIDIA announced DSX Pod, which is a reference design for complete Vera Rubin AI factories built inside modular shipping containers. Partners build, cable, plumb and test the unit in a factory before it ships. On site, it plugs into power, cooling and networking. Source NVIDIA says deployment at a prepared site drops from about six months to as little as 14 days. Jensen Huang said: “An AI factory creates value when it starts producing tokens, not when construction is complete.” Each pod holds seven liquid-cooled Vera Rubin NVL72 racks joined by 17 km of Spectrum-X cable. Seventeen partners were named, including Dell, Foxconn, HPE, Lenovo, Schneider Electric and Vertiv. First deliveries are planned for early 2027. The same day, NVIDIA announced supplier capacity deals. Substrates up 2.5x over three years with IBIDEN, Samsung Electro-Mechanics and Unimicron. Amkor packaging capacity up more than 300% by end of 2028. And a $1.1B advance to KYEC for US test facilities. Jensen said that “capacity is the constraint”in the same release. Furthermore, during his keynote, Jensen Huang talked about a data center math. A 1 GW data center costs about $50B to build and generates about $40B a year in revenue, per his own math. Every month of delay before it produces tokens costs roughly $3.3B. If SpaceX’s CFO is right that compute pays back in under a year, cutting construction time has a direct dollar figure attached to it. [5] Meta Q3 and Muse vs Grok Bot Meta reported Wednesday after close. GAAP EPS came in at $3.83, which includes a $9.89B charge for the settlement with 52 state attorneys general over teen mental health. Without the charge, EPS would’ve been $7.40, beating consensus. Revenue was $64.60B, up 26%, also above estimates. The ad engine kept compounding during this quarter, with ad revenue at $63.06B, up 26% YoY. Most of that was driven by impressions (up 10% YoY) and price per ad (up 14% YoY). Partnership ads passed a $15B annual run rate, up over 97% from a year ago and Threads overtook X in monthly actives for the first time since the launch. Naturally, Muse was the biggest part of the earnings call. Over 10M people in the US and Canada set it up in about six weeks, and about 25% use it on any given day, which Zuckerberg said is more than twice where Threads started. Over a quarter of users shop through it, with an average purchase over $200. Mark mentioned that compute cost is the constraint on a global launch. Because of that, Meta targets more than 10x compute efficiency in the next year. However, the competitive picture is forming since Grok Bot’s daily users are climbing, though its retention hasn’t been published. The biggest difference between Muse and Grok Bot is cost to serve. Both agents burn significant compute per user, and neither has proven it can scale globally without margin compression. Capex came in at $34.91B for the quarter. The full year guide was narrowed to $135B to $145B, pushing the range toward the higher end of prior guidance without raising it outright. Meta didn’t give a 2027 capex number, Susan Li said “we expect to continue investing aggressively” and left it at that. Meta also announced a new nonvoting Class C share, effective Jan 19 2027, which works like a 2-for-1 split. Meta says the shares can be used to raise capital, pay employees in stock and fund acquisitions.. The underlying business beat on every line that was important, but the stock still opened red after being up over 8% when the earnings release came out. [7] Sony and the GTA 6 Trade We first called at roughly $19 a share back in the middle of June and the stock just hit $26. The thesis was simple: Sony has shipped 95.3 million PS5s, the biggest game in history launches on November 19, and Sony’s guidance counts zero from it. We wrote the full breakdown here: Bloomberg reported Thursday that US PlayStation 5 unit sales jumped 22% in September, per Circana data. Hardware revenue rose 43%, driven by price increases over the past year and demand for the $900 PS5 Pro. The PS5 Pro is sold out or scarce at Amazon, Best Buy, MediaMarkt and Yodobashi Camera globally. Twitch chief Dan Clancy has cautioned studios to expect all attention this holiday season to be soaked up by GTA 6. The game hasn’t even launched yet and it’s already moving hardware. Sony’s guidance still treats GTA 6 as if it doesn’t exist and every dollar that flows through the PlayStation Store on November 19 and after sits above that guide. Our co-founder Michael is up over 3x on his Sony calls and the equity is up about 35% since the first article. In a week where every AI name beat and sold off, this was the trade that kept working. [9] Anthropic S-1 The company filed a confidential draft with the SEC back in June and is expected to publicly file its S-1 as soon as Friday. Reuters has reported a raise of up to $100B at roughly a $2T valuation, which would make it the biggest IPO ever. The draft prospectus showed $16.3B of revenue in the first half. The FT separately reported Q3 revenue of $16.8B, up from $11.5B in Q2 and about a fifth of revenue comes from two customers. Cloud compute commitments total about $450B, with roughly $750B including deals still under negotiation. After an investor day, some investors are considering a valuation closer to $1.5T, Reuters reported. Friday’s public filing will be the first audited look at a frontier AI lab’s full economics. [9] Bloom Energy Q3 Bloom Energy beat on everything. Revenue came in at $1.236B, up 138% YoY. Non-GAAP EPS hit $0.94, 40% above consensus. The company raised its full-year guidance for the third time in 2026, now expecting $4.4B to $4.6B in revenue. And even after all that, the stock was down premarket. Product revenue crossed $1B in a single quarter for the first time and operating income for Q3 alone exceeded all of 2025. Revenue from the start of the year is already 50% above last year’s full total and product margin expanded to 38.4% and total gross margin to 34.5%, though installation and electricity margins fell. Last weekend, Microsoft walked away from a planned UK data center deal with Nscale after grid operators said the site couldn’t get power until the 2030s. That’s the problem Bloom solves. Their fuel cells generate power on-site, skip the grid queue entirely, and can be delivering electricity within months of signing a contract. They did it for in under two months last year. They’re doing it now for in New Jersey. As long as grid timelines keep stretching, the order book keeps filling. Bloom Energy Fuel Cells In times like these we think it’s worth watching what companies actually deliver. If the only thing that changed is the price, and the business has a real moat underneath it, you hold. The grid power isn’t getting faster and the demand for on site power isn’t slowing down. That’s why if you have an early entry on , you shouldn’t worry too much about the short term price action. [10] Credo and the MaxLinear AEC Threat Today, opened down 10% after laid out their plan to enter the 1.6T AEC market with Annapurna, a merchant silicon product sold through cable makers. He targeted 20% to 30% share at one or two large US data centers over the next couple of years. In the same breath, he estimated Credo’s current AEC share at 93% to 99%. Credo’s demo cluster at the OCP 2025 However, the two businesses are different up close. Credo designs the silicon, builds the cable, qualifies it with the customer, and books the full product value. MaxLinear plans to sell the chip only, through cable makers. That gives third parties a way to build AEC products without developing their own signal processing, but Annapurna revenue isn’t expected until 2027, with a stronger ramp in 2028. Credo still has relationships with five of the six major hyperscalers and a qualified product shipping at volume today. A 10% selloff on a competitor that won’t ship for a year prices in a lot of damage that hasn’t materialized. We own Credo, and this looks like the market front running a threat that’s still in qualification. [11] What To Watch This week was a lot, but we gathered all the important dates coming up. Make sure to look through them to get a good sense of what’s happening: Oct 23 - Anthropic’s public S-1. Before the open: American Express, SLB, Infosys. Oct 24 (Saturday) - About 2.3B SpaceX shares become tradable. First trading day is Monday, Oct 26. Oct 26 - Amkor Q3. Worth watching after NVIDIA’s packaging deal at GTC Berlin. Oct 27-28 - FOMC meeting. Decision Wednesday. Rates sit at 3.75-4.00% after September’s hike. Oct 28 - SK Hynix. Before the open: GE Vernova, Vertiv, Amphenol. After the close: Microsoft, Alphabet, ServiceNow, KLA. Meta’s next youth trials begin per its 10-Q. Oct 29 - Samsung. September PCE at 8:30am ET. After the close: Intel, Amazon. Nov 2 - Apple fiscal Q4. Nov 3 - Super Micro full results replacing the prelim. AMD. Nov 3-5 - Arm, Coherent, Lumentum. Dates from our Oct 15 notes, not rechecked. Nov 5 - Cook inquiry hearing. Nov 10 - October CPI. [12] Trades We Made Read more
View original →The Laser Shortage Spells Opportunity for This LED Stock
Lumentum’s CEO told Bloomberg this week that the company’s optoelectronic products for AI infrastructure are “effectively sold out through nearly 2029,“ with hyperscalers already giving demand visibility into 2030 and 2031. Lumentum is one of the largest laser suppliers in the world, and it has almost no unallocated capacity for the next three years. Most of that shortage traces back to one material. Nearly every laser used in data center optics is grown on indium phosphide (InP) wafers, and roughly 80% of those wafers come from two suppliers, AXT and Sumitomo Electric. China added InP to its export controls in February 2025, tightening a supply chain that was already behind. A 6-inch wafer now costs about 250% more than it did before the controls, and Goldman Sachs expects the shortage to last through at least 2027. With lasers this scarce, the industry has started looking at links that don’t use them. The most promising approach, usually called “slow and wide,“ replaces one fast laser with hundreds of small, slow emitters running in parallel. The emitters are built on gallium nitride or gallium arsenide rather than InP, and the links can run without the DSP chip that uses much of a transceiver’s power. Microsoft is developing one version with MediaTek for late 2027, and Credo plans to demo a 1.6T cable at OCP next week. In July , Ashkan Seyedi, who was a director of Networking at NVIDIA, left to lead a new AI Photonics unit at ams OSRAM ($AMSS Y / $AMS.SW). ams OSRAM is best known for car lighting, but it is also the largest packaged LED maker in the world and has been shipping microLED arrays in headlamps since 2023. It makes both of the emitters that slow and wide links are built from, microLEDs and microVCSELs, in its own fab. None of this shows up in the share price. The stock is down roughly 85% from its five-year high and trades at 0.7x sales, and seven of the eleven analyst targets we found sit below the current price. Before we begin, this Substack is for informational and entertainment purposes only and does not constitute financial advice. Past performance is not indicative of future results. Never invest more than you can afford to lose. Our writers may hold positions in the securities discussed and may buy or sell them at any time without notice. Table of contents: The Laser Shortage The Four Light Sources Slow and Wide Architecture The NVIDIA Hire The Core Business ams OSRAM’s Moat Risks Valuation [1] The Laser Shortage Inside an AI data center, short connections run over copper and longer ones over fiber, because fiber carries far more bandwidth per cable. Every fiber link has an optical transceiver at each end, and each transceiver holds the lasers that turn data into light. A large cluster needs several of these transceivers for every GPU, Goldman Sachs counts 4 to 6 per GPU on NVIDIA’s newest racks. A 100,000-GPU cluster needs hundreds of thousands of them, and every one needs lasers. As clusters keep getting bigger, the transceiver count scales with them. Demand for transceivers in 2026 and 2027 has grown far faster than forecasters expected. In May 2025, Citi forecast about 56 million 800G and 1.6T units for 2027. BofA’s supply chain checks in September 2026 pointed to as many as 200 million, and Goldman has raised its own 2027 estimate by about 80% since January. The industry is undershipping demand by roughly 30-40% across all major suppliers. Since the start of 2026, nearly every company in the transceiver supply chain has said it can’t make enough and the complaint about supply comes from every layer of the chain, from the substrate makers up to the module vendors: [2] The Four Light Sources Every optical link needs a light source that switches on and off billions of times per second. There are four main ways to build one, and two of them avoid indium phosphide. EML lasers EML lasers are the standard for links of up to about 2 kilometers. Each chip combines a laser and a modulator grown together on InP, so a defect in either part ruins the chip. Because the light leaves from the edge, chips can only be tested after the wafer is cut, and an 800G module uses eight of them. Source: Link-pp CW lasers + silicon photonics Silicon photonics is how most of the industry is responding to the shortage. A CW laser provides steady light while a separate silicon chip, made in standard chip fabs, handles the encoding. One laser can feed several lanes, so an 800G module needs four lasers rather than eight. Silicon photonics made up 38% of AI optical units in 2025 and LightCounting expects 84% by 2030, though the lasers are still made of InP. VCSELs VCSELs are made on gallium arsenide and emit light from the top of the chip, which lets them be tested on the wafer and keeps production cheap. They send light into multimode fiber, where the signal degrades beyond roughly 50 meters. Coherent and Broadcom lead this market, and Coherent is already sampling a 200G-per-lane version. MicroLEDs MicroLEDs are built on gallium nitride, the material used in phone and TV displays. Without a laser cavity , each one is dimmer and much slower than a laser, so links use hundreds of them at once. ams OSRAM works in the last two categories. Alongside microLEDs, it makes microVCSELs, which are VCSELs shrunk down and bonded onto a chip so that hundreds of emitters fire in parallel from one package. [3] Slow and Wide Architecture An EML lane in an 800G module runs at 100 gigabits per second. Microsoft’s MOSAIC research delivers the same 800G over roughly 400 lanes at about 2 gigabits per second each. The industry calls this trade of many slow lanes for a few fast ones “slow and wide.“ The appeal comes down to signal quality and power. At 100 gigabits per second, each light pulse carries two bits using four brightness levels (PAM4), and the levels sit close enough together that small amounts of noise cause errors. Fast modules therefore need a DSP chip to correct the signal, and BofA estimates the DSP accounts for up to half of a transceiver’s power. Suppliers also named DSPs as their tightest component bottleneck in September. At a few gigabits per second, a pulse is simply on or off, and the signal stays clean enough to run without a DSP at all. Power per bit matters because a single GPU cluster already runs millions of optical links. At around 15 picojoules per bit, the energy spent moving data starts to approach the energy spent on computation. Each architecture that moves optics closer to the chip brings that figure down, and Microsoft says MOSAIC uses up to 68% less power than today’s optical links. The limitation is distance. Light from an LED spreads as it travels, and the signal blurs after a few tens of meters. That still covers GPU-to-GPU connections within a rack and between neighboring racks, which is where copper is reaching its speed limits, while longer runs will continue to use lasers. The power saving is the part that matters most because dropping the DSP cuts up to half of a transceiver’s power draw, and across millions of links in a single cluster, that adds up fast in data centers that are already running into power limits. [4] The NVIDIA Hire Ashkan Seyedi was a director of Networking at NVIDIA until July, when he left to build an AI Photonics division at ams OSRAM. Given where he sat in AI networking, he would have had plenty of other options, which makes his choice of a European lighting company worth paying attention to. His team is developing slow and wide optical engines for AI data centers. The work began on the transmit side, which generates the light, and has recently expanded to the receive side, roughly doubling the content ams OSRAM could sell per link. In the September press release, Seyedi described the role this way: “As an integrated electro-optical engine provider, it’s our role to build out the ecosystem required to support a scalable industry solution. ” On the Q2 call, CEO Aldo Kamper said AI Photonics could become a triple-digit million euro opportunity by the end of the decade if the market scales. The technology still has to prove itself but someone who helped shape NVIDIA’s optics roadmap wouldn’t pick this company without reason [5] The Core Business Our valuation rests on the existing business, so it’s worth looking at what that is. ams OSRAM makes chips that emit light (LEDs, lasers, VCSELs) and chips that detect it (ambient light, spectral and time-of-flight sensors), reported across three segments. By end market, automotive was the largest in Q2 at €231M, helped by more LED content in headlamps and sensors. Industrial and medical grew f astest, up 19% to €204M, as horticulture and professional lighting recovered from a long downturn. Consumer contributed €186M, mainly from sensors in premium phones and wearables. TrendForce estimates ams OSRAM held 13% of the $11.7B packaged LED market in 2025, slightly ahead of Nichia at 12%. Its lead in automotive is wider, at 32% of automotive LEDs against Nichia’s 23%, and Omdia ranks it first in automotive light sensors with 31%. Digital Photonics The company groups its newer growth products under the name Digital Photonics. Each of them splits light into many individually controlled pixels. EVIYOS is already selling. The headlamp chip can dim the patch of light that would dazzle an oncoming driver while keeping the rest of the road lit, and ams OSRAM says it has about 80% of the adaptive beam headlight market. EVIY OS had been designed into more than 60 car platforms by Q1, won over €50M of new business in Q2, and is drawing interest from new customers in China. The AR glasses business applies the same approach to full-color microLED displays, built on what the company calls a unique 8-inch microLED platform. ams OSRAM claims its light engine is up to 3x more power efficient than competing designs, with up to 2x higher angular resolution. Omdia forecasts about 60M AI glasses and VR headsets shipping in 2030, which the company translates into roughly 20M glasses with displays at €50 to €100 of ams OSRAM content per pair. UBS expects that content to fall by about two thirds as volumes grow, with revenue starting in H2 2027 and reaching around €100M by 2030. The smallest bet is a time-of-flight sensor that maps depth across up to 48 by 32 zones. In Q2 it won its first designs in household robots, including lawn mowers and four-legged robots, along with its first smartphone design. Restructuring Q2 design wins came to more than €1.6B across over 1,300 projects, bringing the first half to a record of about €2.5B. That total includes a low triple-digit million euro amount of manufacturing work for businesses the company has sold. Those sales are part of a broader cleanup. The specialty lamps business went to Ushio in March and the non-optical sensors went to Infineon on July 1 for €570M, while the CMOS image sensor sale to Indie Semiconductor is signed but hasn’t closed. Together the disposals should raise about €700M, of which about €660M has already arrived. The Simplify program announced in February cuts about 2,000 jobs and targets €200M in annual savings by the end of 2028, about half of it by the end of 2027. Without a clear plan to cut the debt and simplify the business, this would be hard to own. The plan exists, and so far the company is executing on it. [6] ams OSRAM’s Moat Most of the companies working on slow and wide links design the link and buy emitters from someone else but ams OSRAM makes its own. The company already produces microLED arrays at scale. Its EVIYOS headlamp chip packs 25,600 individually controlled pixels and has been shipping since 2023. Microsoft and Credo are designing slow and wide links, but neither has named who makes its emitters. Building that capability takes years, and ams OSRAM says it already runs an 8-inch microLED line in Regensburg. Data link microLED wafer under test, ams OSRAM It also makes both candidate emitters, microLEDs and microVCSELs, which matters while the industry is still deciding between them. Its Regensburg fab grows gallium nitride for microLEDs and gallium arsenide for VCSELs and handles chip processing, packaging and testing on site. The CMOS fab in Premstätten, Austria, stayed after the sensor divestitures, so the company can supply the driver chips alongside the emitters. ams OSRAM showed both at ECOC three weeks ago. Its microVCSEL lanes run about 10x faster than its microLEDs using simple on-off signaling, and the company reports no failures after more than 2,000 hours of stress testing. [7] Risks Reach is unproven No one has publicly demonstrated a slow and wide link much beyond 10 meters. Microsoft and Credo both cite 50 meters but haven’t shown it, and Goldman puts microLED reach below 20 meters with a readiness rating of “Low .“ The underlying problem is that LED light covers a wide range of wavelengths, which causes dispersion and blurs the signal over distance. Hyperlume, now owned by Credo, acknowledges this in its own patent. The architecture may not be adopted BofA expects scale-up CPO switches to be dominated by NVIDIA‘s NVLink at 400G per lane, which is a fast-lane design. Copper remains the default inside the rack, and an ams OSRAM manager on the Optica tour described the ecosystem of connectors and fiber around slow and wide as still evolving. If the approach does succeed, cable makers will hold the hyperscaler relationships, and Coherent and Broadcom, which already produce VCSELs at scale, could move into dense arrays themselves. Even if slow and wide wins, cable makers own the hyperscaler relationship. Coherent and Broadcom already make VCSELs at huge scale and could move into dense arrays. Customer concentration Much of the near-term growth expectation rests on AR glasses, a business that depends on a single unnamed customer. ams OSRAM lost the lead customer for its microLED display program without warning in February 2024 and flagged €600M to €900M of non-cash impairments as a result. J.P. Morgan also reports investor concern about privacy regulation and pricing in smart glasses. Debt Net debt is about €1.74B against €533M of EBITDA, and the sale and leaseback on the Kulim 2 plant adds another €457M. However, the balance sheet did improve this year. In May the company sold €1B of notes due 2032 at 7.25%, with demand at six times the offering, and used the money to retire its 12.25% dollar notes and part of its 10.5% euro notes, saving about €40M a year in interest. Net debt was 3.3x adjusted EBITDA at the end of 2025, or 2.5x after the disposals, and management targets below 2x by 2030. A large payment is still ahead. Minority shareholders in OSRAM Licht AG can sell their shares to ams OSRAM, a €479M liability expected in H1 2027 that is under court review and could rise. If the semiconductor business stalls, there won’t be much cash left after interest to bring the debt down. [8] Valuation Our valuation section is available to paid subscribers only . Join hundreds of asymmetry seekers getting access to: Multiple deep dive articles weekly VIP discord server with private chat and daily trade alerts/analysis from 4 analysts Around the clock support/guidance from Michael, Kawz, and AB analyst team Get Michael and Kawz’s $60M Autopilot fund copy trade updates ($400/year separate value included) Read more
View original →$CRDO: The AI Buildout Comes Down To The Wire
Credo dominates AECs with 88% share and 5 of 6 hyperscalers as customers.
View original →