Neutral
AVGOBroadcomBroadcom
AMD
AMZNAmazonAmazon
GOOGLAlphabetAlphabet
MSFTMicrosoftMicrosoft Research|ANT: S-1 Review and First-Cut EPS Model
Anthropic filed its S-1 publicly today. This note walks through the filing, and at the end we include the Anthropic financial model we built off it. The model rests only on the S-1 and prior public information, so a number of inputs are still open and we will keep refining it. Please contact Sales@Funda.ai to learn more about the EPS model. What we focused on: Revenue and margins: 2Q26 revenue of $11.6bn at a 19% non-GAAP operating margin; 11% for the first nine months How and when ~$800bn of compute commitments flow into costs Whether the cash covers it: ~$132bn on the balance sheet; with IPO proceeds, the model trough is ~$150bn Demand and customer mix: NRR of ~900%; top five customers ~25% of revenue What it means for the AI narrative, and the numbers to watch before the IPO How we built the EPS model I. Revenue 1. Quarterly Revenue Quarterly revenue went $376M, $787M, $1.38bn, $2.04bn, $4.73bn and $11.56bn from 1Q25 to 2Q26. 2Q26 was up 1,368% y/y, while compute and infrastructure expense grew only 300%. Source: Anthropic S-1, FUNDA analysis 2. 3Q is blank, but you can back into it The S-1 leaves 3Q26 revenue blank, but elsewhere states that 3Q26 revenue was 86x 4Q24. With 4Q24 at $197M, that implies ~$16.9bn for 3Q26, in line with the $16.8bn disclosed at the October investor day. Two other data points get you to the same place: TTM revenue up ~1,200% y/y, and TTM revenue per employee above $6.6M on ~5,300 employees. 3. Almost all revenue comes from new models Model versions released in the six months to September 30 drove more than 80% of 3Q revenue. Older models get replaced very quickly. 4. Consumption is the bulk of it Consumption-based revenue was 75% of the total in 2024, 83% in 2025 and 80% in 1H26; the rest is subscriptions. Source: Anthropic S-1, FUNDA analysis II. Run-rate revenue and revenue recognition 1. How run-rate is defined The S-1 defines Run-Rate Revenue as trailing-28-day consumption revenue x 13 plus monthly subscription revenue x 12. The company attaches three caveats: it includes customers later found to have violated the usage policy; it is seasonal, with usage dipping over the summer and winter holidays; growth is expected to slow. 2. Revenue is booked gross When customers buy Claude through the AWS, GCP or Azure marketplaces, Anthropic is the principal and books the full amount as revenue; the cloud providers’ cut sits in “sales, marketing and partnerships.” Marketplace revenue rose from 11% of the total in 2023 to 32% in 2024, 47% in 2025 and 48% in 1H26. Adjust for this before comparing multiples with companies that report net. Source: Anthropic S-1, FUNDA analysis 3. Distillation revenue is now quantified The S-1 states: “we estimate that revenue from such customers represented approximately 5% to 7% of our revenue for the nine months ended September 30, 2026” “Such customers” are those that violated the usage policy or commercial terms. On ~$33bn of 9M26 revenue, that is roughly $1.7–2.3bn (our estimate). Enforcement was stepped up in August–September, so the share should come down. 4. The holiday effect is real Total customer spend in the two weeks to January 2, 2026 was ~20% below the prior two weeks. Any run-rate struck off the last week of December will be depressed by the holidays. 5. Some numbers are left blank on purpose Run-Rate Revenue as of October 31, 3Q revenue, liquidity at September 30, total compute commitments and the price range are all blank. We expect them in the S-1/A. III. Margins (non-GAAP) 1. Gross margin is in the footnotes From 2025 the P&L presents expenses by nature, so there is no gross profit line. Note 8 still breaks costs out by function, and cost of revenue is “substantially all” compute and infrastructure for serving customer workloads. The implied gross margin: −95% in 2024, 40% in 2025, 12% in 1H25 and 85% in 1H26. 85% flatters the picture: the cloud providers’ revenue share and inference for free users both sit in sales and marketing, not cost of revenue. Source: Anthropic S-1 (Note 8), FUNDA estimates 2. Compute intensity is falling fast Compute and infrastructure expense (serving, training, research and internal use) as a share of revenue: 251% in 1Q25 and 144% in 4Q25; 72% in 1Q26 and 48% in 2Q26. Source: Anthropic S-1, FUNDA analysis 3. 2Q26 cost structure Excluding SBC: compute 48% of revenue, sales, marketing and partnerships 18.5% (including the cloud providers’ share), other operating ~9% and personnel ~5%. SBC was $1.26bn, ~11% of revenue. Source: Anthropic S-1, FUNDA analysis 4. Non-GAAP operating margin Annual: −1,273% in 2023, −722% in 2024 and −106% in 2025; +10% in 1H26 ($1.6bn). The S-1 also gives the quarters: −221%, −129%, −93% and −84% through 2025, then −13% in 1Q26 and +19% in 2Q26. 2Q26 adjusted operating income of $2.24bn is GAAP operating income of $0.97bn plus $1.26bn of SBC. 5. 3Q likely back to ~12% At the October 14 investor meeting, the company put the adjusted operating margin for the year through 3Q at 11%. The Information later corrected this to a nine-month figure, not 3Q alone; Bloomberg said “over 11%.” That implies ~$2.0bn of adjusted operating income in 3Q, a ~12% margin, about 7 points below 2Q. We see 2Q’s 19% as a high-water mark. Revenue rose 144% q/q in 2Q while compute is locked in 12–18 months ahead, so compute fell from 72% to 48% of revenue: revenue simply outran capacity. As new capacity came on in 3Q (Google from April, Amazon from May, SpaceXAI ramping into year-end), plus the pricier short-term compute the S-1 flags, margins settled back around 12%. The sustainable level looks closer to 3Q than 2Q. Source: Anthropic S-1, Bloomberg, The Information, FUNDA estimates 6.Cash flow has turned positive Non-GAAP net margin: −1,115% in 2023, −695% in 2024, −97% in 2025 and +16% in 1H26 ($2.7bn). The adjustments are SBC ($2.2bn in 1H26), fair-value changes on convertible notes, derivatives and SAFEs ($72.4bn), other mark-to-market items ($8.7bn), non-cash charitable contributions and the Bartz settlement. Nearly all of the $80.7bn GAAP net loss is accounting noise on financing instruments. Operating cash flow swung from −$2.6bn in 2025 to +$3.9bn in 1H26. With $0.5bn of capex, simple FCF was ~$3.4bn, or 21% of revenue; the S-1 attributes part of that to working-capital timing. Note also that the ~$161bn Broadcom TPU agreement is a finance lease: principal runs through financing cash flow, not operating cash flow or capex, so FCF will flatter the true cash cost of compute. Source: Anthropic S-1, FUNDA analysis IV. Margins will stay volatile, 1. Why Compute is contracted 12–18 months ahead and starts hitting the P&L as soon as capacity comes online, in large blocks and ahead of the revenue it will support. The S-1 notes that margins expanded sharply y/y over the first nine months, but expects them to compress while large amounts of new capacity come on, and to swing meaningfully even as the underlying economics improve. 2. The path in our model Adjusted operating margin of ~12% in 3Q26, ~−9% in 4Q26 and ~−22% in 1Q27, narrowing from there. On a full-year basis: ~−13% in 2027, ~−4% in 2028, back to ~14% in 2029 and ~26% in 2030. The driver is the wave of contracted capacity landing in 2027–28, with costs running ahead of revenue (see Section VI). 3. Short-term compute is the near-term squeeze When demand exceeds long-term contracted capacity, Anthropic buys short-term compute at a higher unit cost. The S-1 expects to lean on short-term compute more in the near term, which will weigh on margins. 4. Compute could get more expensive The S-1 warns compute costs could rise materially, especially if AI demand accelerates or the ability to train frontier models spreads and competition shifts to securing compute. It lists factors largely outside its control: concentrated chip and memory supply, tight power and data-center capacity, and export controls. 5. The response: “Anthropic Direct” The company is moving from cloud-only contracts to a hybrid model: leasing chips directly and placing them in dedicated facilities run by third parties. Data centers take longer to build but are a small share of total cost, so it locks in the sites first and pushes the much larger accelerator commitments later. Once that capacity is live, cheap long-term compute can replace expensive short-term compute. 6. Pricing to customers keeps falling Opus 5.5 costs ~40% less per task than Opus 5; Fable 5.1 is ~25% cheaper than Fable 5 on standard workloads and up to 45% cheaper on complex coding and agentic tasks. The S-1’s argument is that compute efficiency is what lets it cut prices and hold margins. V. The compute bill 1. Disclosed compute commitments total ~$780–810bn Source: Anthropic S-1 Source: Anthropic S-1, FUNDA analysis 2. The mismatch Revenue is usage-based and not locked in; compute is fixed, multi-year and take-or-pay. The Broadcom leases, the credit agreement and the convertible notes are also cross-defaulted. RPO was $10.8bn at June 30, ~75% of it recognized within 12 months, so cRPO of ~$8.1bn covers less than 20% of 2Q’s ~$46bn annualized revenue. Next year’s revenue depends almost entirely on usage. VI. Cash and funding: can they pay for it? 1. What they raised, and what is left From 2023 through 1H26, Anthropic raised ~$140bn in cash (preferred stock, convertible notes and SAFEs), $110bn of it in 1H26 alone, including the $65.2bn Series H. At June 30 it held $112.4bn of cash and equivalents plus $19.3bn of short-term investments, ~$132bn in total, with another $2.5bn of restricted cash. 2. Undrawn capacity A $15bn revolver (upsized in September, undrawn) and up to $77bn of partner note facilities: Google $20bn, Amazon $15bn (tied to compute-delivery milestones) and Broadcom $42bn (usable only for rent). We assume ~$60bn of IPO proceeds in the model (4Q26). 3. Against the megacaps Cash and liquid investments at June 30 (gross of debt): Alphabet $242bn, Anthropic $132bn, Amazon $123bn, Meta $90bn, Microsoft $77bn, Tencent ~$75bn, Alibaba ~$70bn. Anthropic holds more cash than Amazon, Meta or Microsoft, yet the SpaceXAI contract alone runs ~$37bn a year from December 2026. Source: company filings, FUNDA analysis 4. Operating cash flow matters too Compute gets paid from the business as well as the balance sheet. Excluding SBC, non-compute costs were ~33% of revenue in 2Q. The real question is whether cash plus annual operating cash flow can carry the contracted capacity. The figures below come from our model (Section XIV). 5. Compute cost in our model Contracted capacity cost (new cloud contracts expensed as capacity is delivered, commitments in place at end-2025, Broadcom lease amortization and the related data-center cost, AMD depreciation) is ~$24bn in 2H26 and ~$98.6bn in 2027. To smooth the curve, we assume some long-term contracts are signed early in 2028–29 (~$4bn and ~$14bn). Total compute comes to ~$98.6bn / $152.3bn / $182.0bn / $212.0bn over 2027–30. 6. It ties to the ~$750bn long-term figure The Information reported a company plan of ~$750bn of compute spend through 2030. Our model has ~$678.7bn of compute expense over 2026–30. With Broadcom equipment amortized over five years and AMD hardware depreciated over five, ~$61.4bn and ~$8.4bn are still unexpensed by 2030; adding them back gives ~$748.5bn on a spend basis, essentially in line (~$772.6bn including ~$24.1bn of Broadcom lease interest). About $93.4bn of that rests on assumptions beyond the S-1 contracts. Of the ~$819.1bn in disclosed S-1 commitments, 68% is expensed by 2030 and ~$261.3bn falls after 2030. Source: Anthropic S-1, FUNDA model 7. Base-case cash path Exit run-rate of $85bn at end-2026, T28 of $150bn at end-2027, then 60% / 50% / 40% growth over 2028–30; ~$60bn of IPO proceeds in 4Q26 (our assumption). Source: FUNDA model 8. Stress test Source: FUNDA model Against today’s long-term compute commitments, the base case needs no further funding. But the outcome is highly sensitive to revenue growth and compute pricing, and if Anthropic builds more of its own data centers or buys more of its own hardware, compute spend rises and a further raise could be needed. The company has barely touched the money it has raised so far. Still, the conclusion surprised us. We will take a deeper pass and update as soon as anything changes. VII. How Anthropic gets TPUs, and who books the revenue 1. GOOG and Broadcom One is renting from Google Cloud: at least ~$114.5bn from April 2026 to July 2033, take-or-pay. The other is leasing TPU-based systems from Broadcom under five-year master equipment leases: ~$161.2bn of rent, of which the 2026 Lease Arrangement is ~$125.2bn. The TPUs are leased, not bought, but they are accounted for as finance leases and sit on the balance sheet, and the S-1 mentions an early buyout option. 2. How the rent is paid Rent is paid quarterly once the equipment is ready for service. The leases sit in two special-purpose subsidiaries, guaranteed by Anthropic PBC and secured by the equipment. If needed, Anthropic can fund rent by issuing up to $42bn of convertible notes to Broadcom, settled in Class N shares. 3. The data centers aren’t Anthropic’s The hardware sits in dedicated facilities operated by others, with Fluidstack, for example, providing colocation and services. Anthropic has also given developers guarantees and cost-reimbursement arrangements, with a maximum potential payment of ~$0.8bn at June 30. 4. Who books the revenue (Our view) On the GCP route, Anthropic is a Google Cloud customer: this is Google Cloud revenue and sits in the Cloud backlog ($513.9bn at June 30). On the Broadcom route, the lessor is Broadcom, not Google; a finance lease for the lessee usually means a sales-type lease for the lessor, with revenue recognized upfront when the equipment is ready for service. Our current view is that Google books its TPU sales to Broadcom gross, without netting Broadcom’s cost, though net treatment cannot be ruled out. The S-1 adds nothing further. VIII. Demand 1. Key data points NRR of ~900% (excluding Pro / Max subscriptions); customers spending more than $100K / $1M / $10M a year: 11,900+ / 1,900+ / ~250; no single customer above 10% of 9M26 revenue; the top five were ~25%; 79% of 3Q revenue came from outside the top five, up from 64% in 1Q; Claude Enterprise users who also use Claude Code spend ~20x more than those who do not. Source: Anthropic S-1, FUNDA analysis 2. Anthropic runs on Claude In September, more than 90% of code contributions were written by Claude, and code per engineer was 45x the 2024 level. The list-price value of internal Claude usage grew more than 250x from January to September and has exceeded monthly personnel cost (ex SBC) since April. TTM revenue per employee rose from $1.4M to more than $6.6M. IX. Models and training 1. Pre-training is run like a capital budget Improvements are tested at small scale and have to hold up at scale before they go into a frontier run. The S-1 calls frontier runs “planned, scheduled capital deployments.” 2. Post-training matters more and more With the shift from assistants to agents, most of what makes a model useful, accurate and safe now happens in post-training. Agentic skills such as coding, tool use and operating on a user’s behalf are also trained at this stage, on complete, realistic tasks. 3. Beyond scaling laws Four of the co-founders were lead authors on the original scaling-laws paper. The company says proprietary “compute multipliers” from pre- and post-training research let it improve faster than scaling laws alone would, while inference engineering keeps pushing serving compute down. The risk factors also flag that scaling laws could plateau. 4. What each generation brings Deeper reasoning, longer autonomous work and more complex tasks. On one internal benchmark, the success rate reached ~90% in September 2026, up ~50 points in nine months. Training costs are not disclosed: the S-1 says only that frontier runs take weeks to months, cost at least several million dollars, and have sometimes failed or underperformed. What it does disclose is that price per task has fallen with every release. 5. Returns Revenue has grown far faster than compute expense, and new models earn back most of their investment within a few quarters. The flip side: revenue from older models decays quickly, which forces continuous investment in the newest, most compute-hungry models. R&D was ~60% of 1H26 revenue. 6. Where the data comes from Public web data (crawled pages, scanned books), licensed or custom third-party datasets, and synthetic data generated by its own models. The company says it respects robots.txt, paywalls and login gates, but concedes it may not always succeed. The $1.5bn Bartz settlement was approved in July 2026. A large share of customer traffic is covered by zero-data-retention terms. 7. Constitutional training and safety Claude is trained against a published constitution with four priorities: broadly safe, broadly ethical, compliant with guidelines, and genuinely helpful. One interesting detail: training on constitution-style documents alongside fictional stories about well-aligned AI reduces Claude’s misalignment rate. Safety can also slow the roadmap. Anthropic once delayed pre-training of a new frontier base model by more than a month to harden protections against weight theft. In 2026, during pre-release cyber evaluations, a model accessed real computer systems without authorization; the company paused some evaluations and higher-risk RL environments in response. X. How the founders see the future A founders’ letter is unusual in an IPO filing. A few passages stood out to us. 1. Reinvent rather than defend “We will be constantly seeking to reinvent our business, releasing new classes of products whose aim is to dwarf the economic value of our existing ones.” In plain terms: less energy spent protecting today’s products and business model, more on building new categories. 2. The timeline for viable AI “AI chatbots became viable around 2022, and AI coding agents around 2025; AI agents for general knowledge work and scientific discovery are gradually becoming viable in 2026.” Claude Cowork extends Claude Code’s agent architecture to every knowledge worker. 3. How big the market is Global knowledge work worth more than $40tn in 2025 (company estimate); More than 1bn knowledge workers by 2028 (IDC); $22.5tn of cumulative economic impact from AI by 2031 (IDC); Pharma spend alone above $2tn in 2026. 4. “A country of geniuses in a datacenter” Anthropic expects “powerful AI” to be broadly available within a few years. Its definition: expert-level performance in most fields, the ability to use any interface a human can, to work autonomously for hours to weeks, to coordinate across many instances, and to operate lab equipment and robots. 5. Why the compute bets look so large The founders say forward investments can look irresponsibly large at first glance because AI’s value sits on an exponential curve. Their analogy is pharma R&D, only faster and bigger. The hedge is to trade some upside for downside protection: reselling surplus compute, improving efficiency, raising capital. 6. Revenue they chose to leave on the table Claude carries no ads. Claude Mythos Preview was not broadly released because of cybersecurity risk and later shipped as Fable 5, forgoing meaningful revenue; the company also gave cyber defenders ~$100M of usage credits. If needed, they may slow capability progress for safety, even at a competitive cost. 7. The historical reference they chose From 2000 to 2025, revenue at the world’s ten largest tech companies grew ~38x, roughly 11x the pace of global GDP. The founders argue the exponential in machine cognition will matter far more than Moore’s Law. 8. On investor expectations The founders pass on advice from several entrepreneurs who had been through an IPO: “never let your company be worth a dollar more tha
View original →Neutral
AVGOBroadcomBroadcom
AMD
MRVLMarvellMarvell
MSFTMicrosoftMicrosoft
NVDANVIDIANVIDIA
ORCLOracleOracle Research|Optics: OCP 2026 Points to NPO Adoption and New Demand from Copper Replacement
We attended the 2026 OCP Global Summit in San Jose from October 12 to 15, including keynotes from NVIDIA, Broadcom and AMD, and more than twenty optical-interconnect sessions. We also met with Marvell, Credo, Ciena, Cisco, Amphenol, Lightmatter, MACOM and others. Our week in San Jose yielded four key themes, discussed below. This note concludes with detailed notes for each of the companies we engaged with. Contents 1. Key Themes 2. NVIDIA: a CPO version of NVL576 but keeps NPO on the table 3. NPO: How everyone else responded 4. NPO: What users want 5. Pluggables strike back: XPO and USFP and their impact on NPO and CPO adoption 6. Light sources, fiber and test: the constraints in focus at OCP 7. Standards and the ecosystem: MSAs landing in quick succession; an argument against NPO multi-sourcing? 8. Copper’s evolving role 9. Company notes 1. Key Themes Theme 1: Scale-up is still a copper/optics mix; the debate is between NPO and CPO and how outstanding issues with both are resolved Nobody at OCP said copper is finished. Meta will not replace the copper backplane in scale-up for now, while Broadcom sees in-rack copper staying viable for two more years. The role of optics grows with scale: Astera Labs says pods beyond 144 GPUs need optics, with copper in the rack and optics between racks from 72 to 576 accelerators, and optical only above 576 accelerators. Image 1: Astera Labs’ hybrid copper/optical pod: hybrid from 72 to 576 accelerators, all-optical above 576. Source: FUNDA; Astera Labs slide. Where optics is needed, the debate is about where the optical engine sits: NVIDIA showed it soldered onto the switch package (CPO) while keeping NPO on the table. Meta, Microsoft, Oracle, Alibaba and Huawei lean toward board-mounted, socketed NPO. Alibaba also proposed IPO (Interposer-based optics), a solution in between NPO and CPO. A key issue is serviceability. The NPO camp sells “replaceable” while NVIDIA counters that NPO is assembled in clean rooms with woven pigtail fibers, so failures mean a factory return and not replacement in the field. Microsoft and Oracle conceded that without a pluggable module, a single failure has a wider impact. Whoever brings FIT rates down (failures in time, per billion device-hours) and improves testing gets closer to volume orders. Theme 2: XPO and USFP and their impact on NPO and CPO adoption Neither form factor changes the NPO or CPO picture until 400G per lane silicon is tested and proven. XPO is closer to market. It already runs at 200G per lane, and Google and Microsoft are the hyperscalers most interested in the form factor. USFP at 400G per lane will require DSPs, TIAs and drivers that suppliers do not have yet. Once these components are available, both keep pluggable optics in scale-out for another generation and could slow NPO and CPO there. Scale-up stays NPO versus CPO (see the high-density pluggables section). Theme 3: Micro-VCSEL and its impact on NPO and CPO Micro-VCSELs do not compete with NPO or CPO as they are mainly seen as a replacement for copper, and the first products shown at OCP already replace copper rather than optics. Credo’s ZeroFlap ALC replaces AECs, not pluggables or NPO sockets. Later, as scale-in comes in, micro-VCSELs have the potential to replace the links inside the tray that still use copper, since many slow lanes work better there than a few fast ones. ELSFP with micro-ring is another solution for those links, so Lumentum’s external high-power laser business (see the light sources section) should not be seen as affected either way. We see micro-VCSEL as incremental for Lumentum, which moved early in favor of the technology. We believe it is the unnamed second emitter on Credo’s SRMO slide. The field is more crowded than InP and its lead is smaller, so whatever share it wins is upside on top of the core business. The takeaway is a copper-replacement story that grows the optical market and leaves existing optics untouched. Theme 4: Beyond silicon, fiber, lasers and test are now constraints too We kept hearing the same worry: single-mode fiber will be short, whichever approach wins. Lightmatter expects the shortage in late 2027 to 2028 and says the CPO ramp is gated by test equipment in Taiwan; Astera Labs also pointed to component supply as the constraint. Silicon has not dropped off the list. DSPs, TIAs and drivers for 400G per lane are not yet available from suppliers (see the high-density pluggables section). Lumentum’s CPO capacity is booked into early 2029, as its CEO has said. 2. NVIDIA: a CPO version of NVL576 but keeps NPO on the table Keynote: a CPO version of NVL576 Most of NVIDIA’s day-one keynote went to MGX 1.2, 800VDC and the Vera CPU; optics got the last few minutes but carried the most news: Vera Rubin Ultra NVL576 links 576 GPUs over NVLink through a next-gen NVLink 7 CPO switch tray, targets 10T+ parameter models and arrives “next year.” This was probably the first time NVIDIA announced optical scale-up at such a large venue; at GTC it only came up in a small workshop. Image 2: Vera Rubin Ultra NVL576, with 576 GPUs linked through NVLink 7 CPO switch trays. Source: FUNDA; NVIDIA slide. One thing the keynote made clear is that NVIDIA’s next scale-up domain will be optical. NVL576 spans eight racks and 576 GPUs, far beyond the reach of in-rack copper, and NVIDIA is positioning it for 10T+ parameter models, so cross-rack NVLink has to run over fiber. Where the optical engine sits is still open. The keynote showed a CPO version, with optics soldered onto the NVLink 7 switch tray and the slide labeled “Vera Rubin Ultra NVL576 CPO.” In the NPO session, however, NVIDIA’s own slides said it may build NPO with ODMs and will support NPO through NVLink Fusion, and Kuchta mentioned a soldered-NPO proof of concept. Our earlier channel checks pointed to NPO for NVL576’s cross-rack links. We do not read the keynote as a final decision, only as a sign that CPO remains firmly in contention. We expect the choice to depend on yield and test capacity for co-packaged engines, how failures are handled in the field, and laser supply. Either way, demand for cross-rack optics, fiber and lasers is coming; what remains unsettled is how that value splits between the switch tray and the compute side, and between CPO and NPO engine suppliers. Supply-side conversations leaned toward NPO. The scale-up systems people described for 2027 are NPO. They put CPO’s volume in scale-out instead, in NVIDIA’s Spectrum-X CPO switches, where our checks indicate it is still a small share of scale-out switching. They saw CPO inside scale-up as 2029 at the earliest. That is later than the “next year” on NVIDIA’s slide and later than Coherent’s 27H2 target for scale-up CPO production (see the Coherent company note). We take the gap as one more reason to treat engine placement as undecided. On scale-out, NVIDIA presented CPO-based Spectrum-X Photonics multiplane Ethernet, scaling to 512K GPUs on two tiers, with 2.6 ms hardware fault detection, 40% fewer switches, 5x power efficiency and 10x MTBI; Meta, Microsoft and Oracle are listed adopters. NPO session: “Socketed does not mean field replaceable” The NPO session on the 13th had ten speakers, and NVIDIA’s Daniel Kuchta, Principal Hardware System Architect, went last, opening with the line that a few things about NPO “need to be cleaned up.” It was the most direct critique of NPO we heard all week, in three points: • First , based on NVIDIA’s experience, soldered beats sockets. Soldered and socketed describe how the optical engines are attached and either approach can be used with CPO or NPO. NPO’s claimed replaceability advantage comes with its socketed design. Sockets help in prototyping and manufacturing defect screening, but add cost and area, hurt signal integrity and complicate thermals; soldering is cheapest with the best signal integrity. • Daniel Kuchta said the Spectrum-X CPO switch is in production with all 32 optical engines soldered to the package. NVIDIA’s earlier Quantum InfiniBand CPO switch was socketed, hence “been there, done that.” Image 3: NVIDIA’s socketed-vs-soldered CPO comparison: “CPO socketed does NOT mean field replaceable... same for NPO”. Source: FUNDA; NVIDIA slide. • Second , serviceability is “dubious at best.” Even socketed NPO will likely be clean-room assembled with woven pigtails, so it gets shipped back rather than fixed in the field, “a very expensive repair job.” • Third , multi-sourcing is doubtful too. NPO for density or power he understands; for multiple suppliers, “you’re kidding yourself,” since every prototype differs and there are four MSAs, all brand new. NVIDIA did not shut the door, though: the slide below says it may work with ODMs on NPO and will support NPO through NVLink Fusion, and that CPO, pluggables and NPO will likely coexist in scale-out; Kuchta also has a soldered-NPO proof of concept. Image 4: NVIDIA on NPO: it may work with ODMs on NPO and will support NPO via NVLink Fusion, but questions the multi-supplier and serviceability claims. Source: FUNDA; NVIDIA slide. Pluggables stay too. According to Daniel Kuchta, Octal Small Form Factor Pluggable (OSFP) will still be the workhorse. NVIDIA’s announcement of its collaboration with Cisco on Ultra-Dense Small Form Factor Pluggables (USFP) is discussed in more detail in the high-density pluggables section below, together with eXtra-dense Pluggable Optics (XPO) and their collective impact on CPO and NPO adoption. 3. NPO: How everyone else responded Subscribe now Read more
View original →