Subscribe now

I’ll just show a few charts – followed by a long quote of CEO Jassy. If you have doubts about AI ROI, if you still think that we don’t know yet, that there’s no proof to this pudding, you have to read Jassy’s explanation.

So, I won’t repeat one more time that Cloud growth is accelerating, RPO are booming, margins are expanding…

• Google Cloud revenue up 82% YoY. Annual Run Rate $100bn. Operating Margin 36%.

• Microsoft Azure revenue up 43% YoY. Annual Run Rate $120bn.

• AWS revenue up 37% YoY. Annual Run Rate $160bn. Operating Margin 39%.

“AWS is booming, growing 36.7% year-over-year in Q2—our fastest growth in 18 quarters

AWS AI business exceeded $25 billion annual revenue run rate for, growing triple-digit percentages year-over-year.

chips business exceeded a $25 billion annual revenue run rate for its, growing triple-digit percentages year-over-year. Continued gaining momentum with Trainium, with the two leading AI labs in the world, Anthropic and OpenAI, making multi-year, multi-gigawatt commitments”

“our backlog stands at $496 billion, growing triple digits year-over-year”

CEO Jassy:

“We’re reporting $200.6 billion in revenue, up 20% year-over-year.

Operating income was $27.5 billion, up 43% year-over-year.

AWS is booming right now, I’ll share why we’re enthusiastic about the ROIC equation, even with heavy CapEx the next few years

we plan to invest approximately $200 billion in cash CapEx in 2026, the majority of which to support AI and AWS. we have clear line of sight to strong financial returns. I’ll explain why. There are 2 major parts of the investment, the data centers and the servers and networking equipment that go into them.

These are different capital cycles.

Data center capital is spent starting 2 years before we can put servers into them to start monetizing. Once a data center opens with servers plugged in, we start generating significant revenue right away and then get to monetize these data centers for 30-plus years without having to spend that start-up capital again.

Servers and networking equipment operate on a shorter cycle. We typically purchase these a few months before putting them into service, so we have strong visibility into customer demand before we trigger the spend. If the demand isn’t there, we won’t spend the capital. For servers and networking equipment, on average, it takes a little less than 3 years to break even on that investment. The servers currently have a useful life of at least 5 to 6 years.

And most of our AI capacity these days is being contracted for at least 5-year terms. That means that we’re driving significant free cash flow on the servers and networking equipment in the 2 to 3 years after we break even. It’s also worth noting that AWS has a strong track record of finding ways to extend the useful life of this equipment without sacrificing customer experience.

So for our data centers, which have 30-plus year useful lives, we should get at least 5 to 6 generations of server economics, with subsequent generations after the first having even better overall economics because we don’t have to repeat that upfront data center investment.

This means in the short term, when demand is necessitating so many data centers being built simultaneously in advance of when we can start monetizing them. We’ll spend a lot of CapEx and encounter free cash flow headwinds and until these data centers come online can be monetized and we get a few years into these servers being utilized. But as we get a few years out, the revenue growth outpaces the incremental CapEx growth, which will happen at some point. The resulting revenue, free cash flow and return on invested capital is very compelling.

We’ve done this before in the first era of cloud computing.

We now believe we will spend approximately $220 billion in cash CapEx in 2026. The higher cost of memory pushing this number up from a prior estimate of about $200 billion. But even at that amount, we will still not have enough capacity to meet all the demand we have in 2026. And I believe this dynamic will also be true in 2027 too.

In fact, the demand we already have for 2028 is striking. And remember, enterprises are still very early in using inference at scale in their current production applications.

AWS could become very positively be $1 trillion annual revenue business for us in time with very appealing accompanying free cash flow and return on invested capital.”

AI growth is driving non-AI

“We’re seeing strong growth across both AI and non-AI, what we call core-growth in one is driving growth in the other. Growth in AI drives non-AI because post training reinforcement learning and agent tool use is mostly done on CPUs versus AI accelerators

We see a strong linkage between AI spend and core growth (this means demand for non-AI CPU computing cores). As customers invest in AI, we see a core spending increase in core consumption. We expect this relationship to strengthen over time as more AI workloads move into full-scale production and drive additional demand for our core services”

“technically competent companies are going to build their own foundation models, not the really big frontier models, but smaller models that leverage their proprietary data

companies will construct their own purpose-built agents from the ground up, most will also use turnkey agentic services”

2Q26 above Consensus by a mile

Operating Profit 17% above expectations. Strong margins expansion.

3Q26 guidance just inline

Unless we have another margins beat?

Consensus, Valuations

Consensus could be a bit low on revenue growth, high-teens looks more likely than low-teens. Margins expansion is ok.

Stock trading at 23.4x 2027 EPS, 18x 2028.

• Revenue KRW 171.5tn (about $119bn), +28% QoQ and +130% YoY.

• Operating profit KRW 89.5tn (about $62bn), +56% QoQ and +1,814% YoY.

• Operating margin 52.2%, up 9.4 percentage points QoQ.

• Earnings per share KRW 10,849, +52% QoQ. Third consecutive record quarter.

It is a profit beat and a small revenue miss versus Consensus

• Revenue 2% below

• Operating profit 2% above

Samsung is now 100% a memory company.

• Device Solutions (semiconductor division) delivered 99.7% of operating profit.

• Mobile eXperience and Networks (MX / NW) posted an operating loss. First quarterly operating loss in the history of the phone business. OP Margin -2%.

• Visual Display and Digital Appliances (VD / DA) also posted a small loss. OP Margin -0.1%.

1. The whole company is now one segment: memory

• DS revenue KRW 127.5tn, +56% QoQ, +357% YoY.

• DS operating profit KRW 89.2tn versus KRW 53.7tn in 1Q26 and KRW 0.4tn in 2Q25.

• Memory revenue KRW 120.8tn, +62% QoQ, +471% YoY. All time high quarterly revenue and operating profit for the Memory business.

2. Long term agreements will cover 60% to 70% of capacity

• LTAs (Long Term Agreements) signed with the five largest global data centre customers. Final stage negotiations with five more major AI customers.

• Base term five years, rolling, with an additional year added through annual negotiation.

• Target coverage 60% to 70% of total planned capacity.

• Substantial advance payments are a contractual requirement. Management says about one quarter of the total has already been received.

• Minimum price floors are set for mainstream products. Pricing models are differentiated by customer segment and product category.

• This matters two ways. It converts a cyclical business into a backlog business, and the advance payments will distort free cash flow, which the CFO flagged as an input to the shareholder return policy.

3. Supply shortage guided to deepen through 2028

• Lead time from new fab construction to wafer production is more than three years, so incremental industry supply cannot arrive before 2028.

• Unmet 2026 demand carries into 2027. Management expects the 2027 shortage to be more severe than 2026.

• AI frontier model developers are now approaching Samsung directly, bypassing hyperscalers and neocloud providers who cannot secure capacity.

• Mobile and PC demand is moderating as customers raise end product prices, but server DRAM, enterprise SSD and HBM (High Bandwidth Memory) growth far exceeds that moderation.

4. HBM4 ramp is the 2H swing factor, and the share gap should close

• 3Q26 HBM4 revenue guided to more than triple QoQ.

• HBM4 to account for well over 60% of total HBM revenue in 2H26.

• Management target: HBM market share broadly in line with overall DRAM market share by 2H26. That is a large implied share gain.

• HBM4E samples shipped to major customers in 2Q26.

• 2027 HBM supply agreements already signed with customers.

• Balanced mix policy maintained between HBM and conventional DRAM, independent of the profitability difference.

5. The bonus accrual is the reason the headline is not above KRW 100tn

• No incentive accrual was recognised in 1Q26 because negotiations were open.

• In 2Q26 the company recognised a cumulative first half accrual, a special performance incentive equal to about 10.5% of cumulative first half operating profit.

• Part of that accrual sits in work in progress inventory as manufacturing labour cost, so the 2Q26 income statement expense is lower than the total accrual. The remainder hits cost of sales from 3Q26 as the inventory is sold.

• Media speculations that operating profit would have been above KRW 100tn excluding these charges.

Subscribe now

Management messages (conference call)

• Supply through 2028. Memory: “we believe it will be unlikely to see any significant increase in incremental supply through 2028 ... The supply constraints are expected to become even more severe in 2027 than 2026, reinforcing our view that the supply shortage will persist through 2028.”

• Demand source. “As many AI frontier model developers struggle to secure sufficient cloud capacity from the hyperscalers, they are increasingly turning to the neocloud providers ... the AI frontier model developers increasingly approaching us directly to secure their Memory supply.”

• Demand exceeds supply. “Despite our efforts to increase production, our customers demand is outgrowing our supply.”

• Business model change. “we expect to transition our business structure from one that was previously overexposed to supply-demand cycles to a more stable and predictable model.”

• LTA coverage. “we believe that DRAM and NAND supply volumes under multiyear supply contracts will easily account for 60% to 70% of our planned capacity.”

• LTA advance payments. “we have included substantial advanced payments as a contractual requirement to be maintained over the multiyear contract period ... we have already received one-fourth of these payments.”

• LTA price floors. “For mainstream products, we have established minimum price floors at a level to ... mitigate against future investment risk arising from market price volatility.”

• HBM4 ramp. “expect HBM4 sales increased by more than threefold Q-on-Q in the third quarter ... HBM4 is expected to account for well over 60% of our total HBM revenue mix.”

• No ADR. CFO: “we are not currently reviewing an ADR issuance. That said, from a mid- to long-term shareholder value enhancement perspective, we view it as one of several possible options.”

• Foundry turnaround. Foundry EVP Sukchae Kang: “While it is difficult to predict exactly when we will turn around ... we believe the turnaround is possible in the near term.”

Consensus, Valuation

Consensus looks ok for 2026 and 2027. Too low for 2028. Same as other Memory names, or AI Semis in general. Consensus doesn’t yet dare to forecast 30% growth in 2028. That will come when analysts have to think about that, ie early 2027.

Samsung Elec trades at 4.4x 2026 EPS, 3.1x 2027. That’s close to ridiculous, and really ridiculous if LTAs end the cyclicality of the Memory market.

Subscribe now

1. Strong and accelerating Cloud growth

Numbers are impressive.

Microsoft closed its FY26 in June. For FY26,

• Total revenue $332 bn, up 18% YoY (FY25 15%)

• Microsoft Cloud revenue $214 billion, up 27% (FY25 23%). Nearly 90% of revenues from customers outside of frontier model companies.

• Intelligent Cloud revenue $138bn, up 30% (FY25 21%)

• Azure surpassed $100 billion, up 41%

June-26 (4Q26): Cloud growth accelerating

• Microsoft Cloud growth stable (27% YoY)

• Intelligent Cloud growth accelerating (32% YoY)

• Azure growth accelerating (43% YoY)

2. Backlog is enormous, and the growth is broad beyond OpenAI

• Commercial RPO (Remaining Performance Obligation) $678bn, +84% YoY, up from $627bn in 3Q26.

• All sequential RPO growth came from customers outside the frontier model companies. RPO +25% excluding OpenAI.

• Weighted average duration 2.3 years. Roughly 30% converts to revenue in the next 12 months, that near-term portion increased +37% YoY. The portion beyond 12 months increased +112% YoY.

• Related party disclosure: FY26 revenue from commercial arrangements with OpenAI, including revenue-sharing payments, was $24.1bn. Accounts receivable from OpenAI $6.0bn. Microsoft holds an approximate 25% as-converted equity method interest, with $13.0bn of funding commitments of which $11.9bn funded.

• Microsoft tries to explain that Cloud growth is not driven by OpenAI or Anthropic.

  “All sequential commercial RPO growth was driven by commitments from customers outside of frontier model companies“

3. No more open buffet, “we are moving from seats to a seats plus consumption model”

CEO Nadella repeated 2 times: no more open buffet pricing.

• “we are also evolving our business model beyond per-seat to per-seat-plus-consumption”

• “we are also moving from seats to a seats plus consumption model”

• “stronger-than-expected GitHub Copilot consumption following the June business model change to align pricing with usage and value”

• Model shift from per-seat to per-seat plus consumption across M365, Dynamics 365 and GitHub. Usage-based billing added to Cowork earlier in July.

• Copilot revenue accelerated over 60% quarter-over-quarter.

• Copilot over 30 million paid seats, up from over 20 million in 3Q26.

• GitHub Copilot 50 million users. Usage-based billing introduced in June. GitHub Copilot revenue accelerated over 60% QoQ.

Subscribe now

4. Capex is unchanged but lower… Accounting trick to lower reported capex

• Useful life of datacenters and office buildings extended from 15 to 25 years, effective start of FY27. Minimal benefit to FY27 operating income.

• The real effect is on capex classification. More future datacenter leases become operating leases rather than finance leases. Finance leases are included in capex, operating leases are not.

• Calendar 2026 capex expectation therefore moves from approximately $190bn (April guide) to approximately $175bn, with no change to underlying investment plans.

• 1FY27 capex expected to grow YoY.

• Off balance sheet scale to watch: leases not yet commenced, primarily datacenters, total $329.1bn as of 30 June 2026, commencing between FY27 and FY33 (10-K).

“we are extending the estimated useful lives of our datacenters and office buildings, from 15 to 25 years, reflecting our operating history and expected use of these assets.

“This change affects only the timing of future depreciation and is expected to have a minimal benefit to FY27 operating income.

“The greater impact is on capital expenditures as more of our future datacenter leases will shift from finance leases to operating leases. Finance leases are included in capital expenditures while operating leases are not.

Outside of this useful life impact, our calendar year 2026 CapEx investment expectations remain unchanged. However, the shift from finance to operating leases adjusts our expectation to approximately $175 billion”

5. FY27: Cloud revenue up, Traditional revenues down

First management gave a warning on Traditional revenues that will decline:

• M365 Commercial products (that’s Office 365 online) and Server products (old style Enterprise IT, SQL): we expect revenue to decline in the mid-single digits for the full fiscal year.

• Windows OEM and Devices (that’s Windows in PC) will be impacted by lower PC market: we expect revenue to decline in the high-teens for the fiscal year.

But Commercial & Cloud will remain “strong”:

• At the company level, with strong commercial momentum, we continue to expect another fiscal year of double-digit revenue and operating income growth.

FY27 capital expenditures will grow year-over-year given demand signals across our portfolio.

Even as we invest to meet growing demand, full fiscal year operating margins should be down less than a point. In addition, we expect to remain free cash flow positive in FY27.

4Q26 results (June 2026 quarter)

Small beat

Next quarter guidance (1Q27, September 2026 quarter)

• Revenue $89,850m to $90,950m, growth of 16% to 17%. Consensus into the print was about $89.66bn. Guidance is inline.

• Cost of revenue $29,600m to $29,800m, growth of 23% to 24%.

• Operating expenses $16,800m to $16,900m, growth of 7% to 8%.

• Operating margins relatively flat YoY.

• Productivity and Business Processes revenue $36.85 bn, or growth of 12% YoY

• Intelligent Cloud, we expect revenue of $41.1 bn, or growth of 33% YoY

• In Azure, we expect revenue growth of approximately 45%

• More Personal Computing, we expect revenue to be $12.4, or decline of -9% YoY.

Subscribe now

Valuations

Main points

1. 2Q26 is a miss, by popular vote (Korean media, stock price down -10% in Korea)

• as Operating Profit was -5% below Consensus. DRAM and NAND bit growth was inline with Guidance. The miss was because DRAM ASP increased less than expected, less than Commodity DRAM.

• Why? Maybe a timing issue (HBM4 ramping, tech conversions). Maybe a mix issue (SKH has a bigger share in HBM ~58%, hence a smaller exposure to Commodity). Maybe it’s structural as AI related revenues are covered by Long Term Agreements, which anchor prices.

• The point of LTA is to provide volume visibility to both buyer and seller, and also a pricing mechanism which lowers the volatility of pricing – hence decoupling from volatile Commodity prices. Note that under LTAs Pricing is not fixed and not uniform but “designed to address price volatility”.

• So that’s the paradox. The stock price reaction in Korea reflects the belief that Memory is still a volatile / cyclical industry – and a small miss means cycle-peak and upcoming collapse – and people sell the stock. But it could be the opposite: the miss is due to LTAs which lower pricing volatile and also support capex / volume growth and a reduced cyclicality means that the stock should be trading at 15x, not 5x.

Net profit KRW 93.9tn is inflated by a 63tn won investment asset gain.

2. The miss is a mix and timing issue, not a demand issue

• DRAM bit growth was high-single-digit % QoQ, in line with the guidance given at 1Q26. NAND bit growth was mid-teen % QoQ, also in line.

• DRAM ASP rose about 30% QoQ. NAND ASP rose mid-50% QoQ. The DRAM number is what disappointed.

• Management explanation: “In the second quarter, shipments of some high value add products were pushed back into the second half, and changes in the product portfolio appeared to have affected our blended ASP.”

• In the background of the “lower” ASP increase, SK Hynix now has signed Long-Term Agreements LTAs with ~10 customers. ~5-year terms with deposits and volatility-adjusted (not fixed) pricing. This structurally caps spot upside, which is part of why a spot-price-driven consensus was missed.

• Cost side was excellent. Cost of goods sold rose only 22% QoQ on revenue up 51%. Gross margin 83%.

3. HBM4, 1-c DRAM node and 321-layer NAND are in mass production and the 2H earnings driver

• Mass production shipments began in 2Q26. Full ramp planned for 2H26.

• Yield and quality “today nearing the levels of HBM3E, which is already in the maturity stage”.

• “These factors are likely to gradually ease in the second half. As HBM4 shipments ramp up in earnest and the 1c-nanometer conventional DRAM shipment increases, we expect bit growth in the second half to be higher than the level in the first half. In addition, considering customer demand and the changing product mix, the growing HBM4 sales and the higher contribution from value-add products will have a positive impact on our blended ASP as well.”

• “In the previous quarter, our 321-layer product accounted for the largest proportion of NAND production, and we plan to expand the share within our domestic capacity to the 50% level by the end of the year as planned”

SK Hynix does not provide a complete guidance. Only bit growth for the next quarter.

Which means that you only need to get a feeling for price movement in 3Q. Refer to industry participants, specialized websites, etc

For 3Q26, it suggests (based on my estimates) that results will be ~7% above Consensus as it was 2 weeks ago.

4. Long-Term Agreements (LTAs) with around 10 customers change the earnings profile

• LTAs concluded with around 10 customers including key strategic partners. Discussions continue with other major players.

• Typical term around five years. Pricing is not fixed and not uniform, designed to “address price volatility”.

• Deposits and other financial mechanisms included to enforce purchase commitments and improve demand visibility.

• Management will not disclose what share of sales is covered, but frames the objective as “downside resilience” while keeping flexibility to capture upside.

• 2027 HBM volume and price negotiations are underway. Management says rising conventional DRAM prices “may also have some influence on our HBM pricing discussions”.

• This is a structural change: it caps some upside in a spot price spike, which is part of why the quarter missed a spot-price-driven consensus.

• From TrendForce

5. Capex raised again and the balance sheet is now a weapon

• 2026 Capex guided to the high-40tn won range (roughly $33bn). At 1Q26 the guidance was only “increase significantly compared with last year”.

• 2025 Capex was KRW 28.5tn

• 2024 Capex was KRW 16.7tn

• M15X mass production pulled forward. Yongin Fab1 cleanroom opens early 2027. P&T7 advanced packaging, M17 NAND base and a new domestic cluster to be phased.

Management messages

• Demand base. President Song Hyun-jong: “we are witnessing a structural shift in demand where both AI memory and conventional memory are growing together.”

• Customers want more. “our major customers are still requesting more memory supply.”

• Supply. “it appears difficult for the supply-demand balance to improve meaningfully in the near term”, citing process complexity for HBM and AI server memory plus fab construction lead times.

• 2H acceleration. “As HBM4 shipments ramp up in earnest and the 1c-nanometer conventional DRAM shipment increases, we expect bit growth in the second half to be higher than the level in the first half.”

• ASP recovery. “the growing HBM4 sales and the higher contribution from value-add products will have a positive impact on our blended ASP as well.”

• HBM4 quality. “we began mass production of HBM4 for key customers in the second quarter, with the yield and quality today nearing the levels of HBM3E.”

• Smaller AI models are not a threat. On more efficient AI models: “we do not see these more efficient AI models reducing infrastructure demand. Rather, we believe that they will further broaden the AI base.”

• On Hyperscalers capex fears. Data centre leasing and efficient models are “not as a sign of AI investment slowdown, but more as a transition towards higher utilization of the AI infrastructure that has already been built at scale.”

• Capex. “Due to the schedule acceleration and investment expansion, our 2026 CapEx is expected to reach the high-KRW40 trillion-range.”

• Supply as competitive advantage. “in the AI era, technological leadership alone is not enough. The ability to supply the required volume at the right time has also become a critical part of competitiveness.”

• Shareholder returns. “we intend to communicate our plans to the market within the year once they have been finalized.”

• HBM4E: 12-layer samples shipped to major customers in 1H26 (announced 18 June 2026), 48GB, 16Gbps per pin, power efficiency up more than 20%, heat resistance improved 17% versus HBM4 using Advanced MR-MUF. Volume production targeted 2027.

• HBM5: developing iHBM, integrating cooling elements inside the package, expected to cut thermal resistance by more than 30%.

From Counter Point:

Subscribe now

2Q26 results

Volume and pricing

• DRAM bit growth high-single-digit % QoQ, in line with guidance. ASP about +30% QoQ.

• NAND bit growth mid-teen % QoQ, in line with guidance. ASP about +50% or better QoQ (mid-50%).

• Enterprise SSD revenue doubled QoQ. Solidigm revenue from 30TB and above high-capacity eSSD more than tripled QoQ.

• SOCAMM2 sales grew significantly. 1c-nm (10nm-class 6th generation) shipments began in earnest.

• 321-layer NAND is already the largest share of NAND production, targeting about 50% of domestic capacity by end-2026.

Net profit is inflated by a one-off. Use operating profit.

• Net profit 93.92tn won includes net non-operating profit of 62.17tn won.

• That comprises FX-related gains of 1.15tn won and other non-operating profit of 60.89tn won, which includes investment asset related gains of 63.27tn won.

• Pre-tax profit of 122.71tn won is an all-time KOSPI record, previously 58.83tn won by Samsung Electronics in 1Q26.

• Reported basic EPS of 132,126 won is therefore not a run-rate number.

• Debt-to-equity ratio 7%, down from 12% at 1Q26.

• Working capital build of about 32.2tn won in the quarter, mostly receivables. Worth monitoring.

• Capex of 10.7tn won in 2Q26 against full-year guidance in the high-40tn won range implies a heavy 2H.

3Q26 guidance

• SK Hynix does not guide revenue, margin or EPS. It guides bit shipment growth and capex.

• DRAM bit shipments approximately +10% QoQ, with an active response centred on server products.

• NAND bit shipments low-single-digit % QoQ, including Solidigm, driven by 321-layer products and SSD.

2H26 bit growth expected to be higher than 1H26. Blended ASP expected to improve on HBM4 ramp and higher value-add mix. Management expects “the pace of earnings improvement in the second half to strengthen further”.

This is my best guess. I add 4Q26 as well. Numbers in yellow cells are my estimates.

For 3Q26, it suggests (based on my estimates) that results will be ~7% above Consensus as it was 2 weeks ago.

Conference call highlights

• CSP capex fears. Data centre leasing and efficient models are read as higher utilisation and faster monetisation, not a slowdown. “We do believe AI infrastructure investment will remain solid beyond next year.” Customers are still asking for more supply.

• Oversupply risk. Capacity plans are anchored on LTAs and customer discussions, executed in phases. “we do not believe our medium to long-term investment plans will lead to oversupply right away.”

• LTA structure. Term normally around five years. Pricing mechanisms vary and are designed for volatility, not fixed price. Deposits included. Share of sales covered will be kept “at an appropriate level” and is not disclosed.

• HBM4 competitiveness. Yields and quality near HBM3E maturity. Focus now is capacity ramp. HBM4E samples complete, volume from 2027. iHBM in development for HBM5, cutting thermal resistance by more than 30%.

• 2027 HBM pricing. Negotiations “progressing smoothly”. Conventional DRAM strength may influence HBM pricing but HBM pricing is not set by conventional DRAM alone. Inputs include wafer consumption, TSV and packaging capacity, technical complexity and customer value.

• Capacity location. Icheon and Yongin remain the DRAM and AI memory hubs, Cheongju for NAND and advanced packaging. Domestic versus overseas will be decided on power, water, people, supply chain and customer access. “at this time, there are no further decisions made in addition to the investment that has already been announced.”

• NAND strategy. Not a choice between SLC, TLC and QLC. High-capacity QLC eSSD for AI data lake and HDD replacement, high-performance TLC eSSD, and SLC-mode SSD. New tier products for KV cache offloading and near-GPU storage.

• ADR mechanics. ADRs freely convertible into Korean shares from 30 July. Conversion of shares into ADRs is limited and slow. Conversion limit set at 17,790,000 shares. No decision on increasing the ADR proportion.

• Capital allocation. Three goals: timely growth investment, sound financial position, shareholder returns. Additional returns under evaluation but ADR offering rules prevent disclosure of new material information. Plans to be communicated within the year.

Subscribe now

Consensus, Valuations

What matters most

AI Computing demand continues to shift to CPU

• We know since 2 quarters ago that growth in Inference and Agentic compute is driving rapid growth for CPU “AI expands from training to inference; and increasingly to agentic and multi-agent systems, general-purpose server CPU density continues to increase, and our core server CPU franchise is growing faster than ever”

• Intel is capacity constrained “our top priority is to ramp capacity and improve factory output as quickly as possible”

• Demand outlook is increasing “outlook for server CPU demand has improved again since our last earnings report, and we’re forecasting strong double-digit unit growth for the industry this year and next, with momentum extending into 2028”

Supply is the constraint; Intel is increasing Capex

• AI demand is outstripping supply across logic, memory and substrates

• Intel’s Foundry capacity increased is “more skewed to the end of Q3 into Q4”, i.e. supply will increase more in 2H26

• 2026 Capex raised to more than $20bn, 2027 Capex guided significantly higher “due to strong customer demand signals, we are raising our outlook for 2026 and now expect our Capex to be more than $20 billion, which is up significantly versus our expectations entering the year”

• EUR5bn investment to expand Intel 3 Xeon capacity in Ireland; US tools and space spend 2021-2026 approaching $100bn.

Management messages

• Demand vs supply - CEO Lip-Bu Tan: “strong demand for our products continues to outpace our growing supply... Today, we are seeing the strongest revenue growth in more than fifteen years.”

• Three strategic assets - Tan: Intel is positioned around “our x86 CPU franchise, our advanced packaging technology, and our vast wafer foundry network.”

• Foundry confidence - Tan: “my confidence in our foundry process roadmap has grown significantly... 18A output increased meaningfully in the quarter; yields continue to track ahead of expectations.”

• “our factories across Intel 7, Intel 3, and Intel 18A exceeded internal volume targets, driven by improving yields, better cycle times, and increasing wafer-starts”

• 14A committed - Tan: “we remain on track for 14A risk production for our internal products in the second half of 2027 and we made the decision in Q2 to fully commit to high volume ramps in 2028.”

• Beat drivers - CFO Dave Zinsner: revenue was “$1.8 billion above the midpoint of our guide”; AI-driven businesses “grew greater than 70 percent year over year... and contributed approximately 70 percent of revenue.”

• Gross margin - Zinsner: 41.8%, “approximately 280 basis points better than guidance... driven by higher revenue, better yields, and higher ASPs due to mix and pricing actions.”

• Margin goal - Zinsner: the number one goal this year was “to get gross margins comfortably into the 40s in every quarter.”

• CapEx - Zinsner: 2026 CapEx “more than $20 billion”; 2027 “significantly above the 2026 levels with the vast majority spent across our US network.”

more below the paywall

Subscribe now

Read more

Cloud revenue up 82% YoY, now at a run-rate of $100bn/yr. Cloud Operating Profit up 210% YoY, on Operating Margin 36%. A year ago, revenue run-rate was 55bn/yr, Operating Margin 21%. Management says “Google Cloud acceleration in growth led by enterprise AI Solutions and enterprise AI Infrastructure”. Meaning: Gemini Enterprise, Gemini Cybersecurity and also Anthropic.

Cloud revenue Backlog (Remaining Performance Obligations) reaching $514bn, up 375% YoY. That’s 5 years of Cloud revenues at the current run-rate. If you think that Cloud growth will slowdown, or that Capex will not increase – think again.

We’re seeing a stunning acceleration in the growth of Cloud revenues and Cloud Backlog, which support a high expansion in margins.

2026 Capex guidance revised up to $200bn from previously $185bn. CFO “continue to expect our Capex to increase significantly in 2027”. This level of Capex leads to negative Free Cash Flow, or the 3^rd quarter of FCF hovering around zero. You can expect that to continue in 2H26 and 2027.

Total revenue up 24% YoY, Operating Profits up 30% YoY. This doesn’t beat expectations: Operating Profit is inline with Consensus expectations. Core revenues (Search, YouTube, etc) up 15% YoY (stable) but with a little blip in Operating Margins down to 43%. Look at chart below, there’s some mild volatility in OPM, along a trend line that’s clearly going up.

Don’t look at Net Income or EPS as there is a massive non-ops equity gain of $98bn (that’s SpaceX and Anthropic) when Operating Profit is “only” $41bn.

2Q26 Operating Profit is inline, Capex increase, negative Free Cash Flow and the stock is down -2% after-hours. The stock is not expensive at 25x forward EPS for 30% YoY growth in Operating Profits in 1H26 (let’s forget 1H26 Net Income up 180% since that’s due to investment gains).

More below the paywall

Subscribe now

Read more

Subscribe now

Management messages

• AI demand - CEO C.C. Wei: “AI-related demand continues to be extremely robust... our conviction in the multi-year AI megatrend remains very high.” FY26 revenue growth now seen slightly above 40% in USD.

• Duration - Wei: “I believe from this day on all the way to probably 2029, 2030, the demand is very strong... we are witnessing a kind of a new industry... called AI industry.”

• AI CAGR: the January guidance of mid-to-high-50s AI accelerator revenue CAGR is now “stronger and stronger and stronger” - no updated number because it keeps increasing.

• Agentic AI - Wei: “The emergence of agentic AI is leading to a resurgence in the role of CPUs in AI data centers... whether it is x86, ARM-based, or RISC-V architecture, they are almost all TSMC customers.”

• Supply and demand: leading-edge and advanced packaging capacity remain in shortage; management is checking AI data center buildout progress to make sure chips do not end up in inventory.

• Demand is very strong through 2029-2030. Asked if unconstrained 3nm-and-below demand exceeds supply by more than 30-50%, Wei: “the gap is very big,” no number shared. 2nm family capacity CAGR shown at the symposium (about 70% for 2026-28) is “now bigger.”

• Mature nodes tight only in power management ICs and sensors; other consumer-driven mature demand is weak.

• Pricing and margin: TSMC will not squeeze customers with 4-5x price hikes (that’s a reference to Memory prices). On memory makers 86% gross margin: “68%, I will be happy about that.”

• CapEx - CFO Wendell Huang: budget raised to US$60-64bn; “the CapEx in the next three years will be even more significantly higher than the past three years.” No capacity bottlenecks foreseen.

• Arizona - Wei announced the additional Arizona investment for “several more semiconductor logic wafer fabs for 2-nanometer and below technologies as well as advanced packaging fabs.” (Q&A and press confirm US$100bn additional, US$265bn total.)

• Taiwan and Japan fabs also being accelerated.

• Competition - Wei: “no shortcut” in foundry, “the most important thing... is the technology, manufacturing, and customer trust... choosing a technology, ramping it up, it is not buying milk from 7-Eleven.”

• TSMC welcomes Intel EMIB-T advanced packaging as relief because TSMC backend capacity shortage is limiting customer growth; backend competition does not open a path into frontend wafer share.

• Technology: A14 on track, pre-production 2027, volume 2028. A13 and A12 derivatives in 2029. Adoption of High-NA EUV still edpends on cost and maturity, cost problem is that field size is half that of Low-NA. COUPE (co-packaged optics) now in production and ramping, important over the next few years.

• Dividends - Huang: NT$24 per share in 2026, +33% YoY (NT$18 in 2025), with a further increase expected in 2027.

Below the paywall:

2Q26: 5% above Consensus

3Q26 guidance: 6% above Consensus

Reasonable forecasts to 2028

Consensus, Valuations

On valuations chart above: 1) Consensus is a bit too low, say 10% too low so forward PEx closer to 18x or +1 standard deviation 2) if you agree on the AI “mega trend”, we’re looking at next 2-3 years at ~30% growth 3) is the stock reaching its upper PEx or how much re-rating is reasonable given TSMC own growth and peers’ valuations.

Read more

Main points:

• 2027 low-NA EUV capacity is “close to fully covered with orders”

• adding about 30% capacity for EUV in 2027 versus 2026.

• we strongly investigate another 30% increase in our EUV capacity for 2028.

• DUV grows for 2027 and for 2028, we are going to look into a 30% increase of our capacity for both years.

• 30% more tools... if you recognize the difference in outputs, then in essence, what you’re looking at is not 30% improvement of wafer capacity that we’re adding, but approximately 45%.

• Customer demand remains very strong with visibility now extending several years into the future.

• The dynamics are similar in advanced logic and DRAM, the plans to build up capacity are equally aggressive.

• Our customers in both segments are entering into long-term agreements with their customers, providing them with longer-term visibility and the confidence to add significant capacity to support demand.

• Logic: continued investment expansion of 3-nanometer capacity in support of the latest generation of AI accelerators. Also 5-nanometer and the 4-nanometer nodes to support the diverse set of chips required by AI products. 2-nanometer node continues to ramp rapidly. Already planning investment to support 1.4-nanometer nodes. These dynamics in the logic segments are driving both an increase in litho intensity and greater demand for advanced lithography.

• DRAM: the supply challenges driving up both DDR and HBM prices have prompted significant investments in fab expansion. Our customers are adding meaningful capacity this year, while at the same time, plan further capacity expansion, multiple megafabs. DRAM lithography intensity is raising as customers migrate to advanced nodes. This includes both EUV and Deep UV immersion.

Below the paywall:

2Q26: a little ~10% beat

3Q26: Operating Income guidance a massive 35% above Consensus

Simple forecasts to 2028

Consensus, Valuations

Subscribe now

Read more

Context

Since ChatGPT hit the masses in 2022, we’ve had multiple scares and apocalypse theories.

• LLM are commodities that can’t monetize,

• Then, LLM can monetize but there’s no ROI,

• Then, Capex is a ponzi scheme of circular financing,

• Then, GPU are obsolete in 1 year, depreciation over 5 years is a scam,

• Then, Agentic AI costs too much money and firms are cutting down,

• I forgot several episodes of China has better, cheaper models (DeepSeek),

What’s happening today?

• Hyperscalers Free Cash Flows (FCF) are going to zero and therefore, they’ll have to cut down AI Capex? (I told you in March-1, March-2, April, that FCF are declining but growth + margins increasing – the real problem is asset turns declining)

• There’s another theory that says that Semi margins are too high (Micron 80% operating margin, Nvidia 65%), squeezing profits of App companies. Margins must decline?

• Another theory is “negative externalities”, AI Data Centers are creating too much inflation (electricity, land, memory, etc) and nuisances (noise, heat, water), and negative social impact (AI destroys jobs), so the Data Center build out will slowdown.

Of course, these theories catalyze after a strong stock rally, which sustains the narrative “the bubble ready to burst”. We are here again. We’re at a juncture of a bubble narrative (strong stock rally), and a peak AI theory (peak capex, AI is flawed in some ways).

After this long intro, what I got for you below is:

• AI Semi and Hyperscalers valuations: stocks are not expensive.

• Have you looked at Hyperscalers’ future Cloud revenues, ie Remaining Performance Obligations (RPO)?

Semi and Hyperscalers valuations: stocks are not expensive

Since April, AI Semiconductors have rallied, and Hyperscalers have drifted down.

Semis aren’t expensive, trading at average forward PEx 17.7x

This big rally is the result of massive earnings growth. I’ve read that too many times, the 1Q26 earnings of AI-related Semi firms were stunning. Micron 2Q is even better. Let’s wait 3 weeks to get ASML, TSMC, Intel, AMD.

There’s at least 2 things that you can worry about:

• The runway: high growth for how much longer? Does it stop in 2028? Or revenue decline in 2028?

• The extortion: net margins have doubled in 5 years, from 24% to 48%. Semi vendors are a lot more profitable than their clients (Hyperscalers). Can this be sustainable?

Hyperscalers are cheapish at -1 standard deviation or forward PEx 21.1x

Consensus is reasonably forecasting steady revenue growth (16-18% does not look stretched) and stable margins (24-25%). A little margins hit in 2027 as D&A increases. That looks fine.

A lot of people worry about zero / negative free cash flows, fund raising (debt / equity), cancellation of dividends and buybacks. Indeed, declining Free Cash Flows can freak people out. Sum of Alphabet, Amazon, Meta, Microsoft below:

This goes together with declining Net Cash positions. All hyperscalers are close to zero net cash.

Maybe that’s a limiting factor to AI Capex growth in 2027-29. But again (refer to report links above), Cloud revenue growth is accelerating, margins are increasing. And maybe Hyperscalers can have a little bit of leverage? Does it matter to have 1x Net Debt / Ebitda?

1 year of Ebitda = $700bn and as you can see below, Ebitda-$ is increasing.

Hyperscalers have started to issue debt. I think we should expect more of that.

The market is missing this!!!

I receive lots of emails titled “The market is missing something”. I hate that line. The market is a weighting machine of millions of opinions. It doesn’t miss anything. But one opinion dominates price action. It’s often a short-term story that pretends to be a long-term view: AI will never be profitable – that the story of today. Tomorrow that will change.

I have a theory for you. Have you looked at Hyperscalers’ future Cloud revenues, ie Remaining Performance Obligations (RPO)?

I think the main freaking out factor is composed of 2 things:

• There is profit taking, some rotation, some market reaction to inflation / rates: stocks go down a bit. IMO, that’s fine.

• The more fundamental concern is (still) AI’s future growth and profitability, ie can the high AI Capex continue or we’re gonna hit an economic wall? 4 charts below. No conclusion but think about this:

As we know, AI Capex (by hyperscalers) has gone ballistic:

As we know, Cloud revenues are accelerating from ~22% YoY in 2024, to ~35% now:

Future Cloud revenues,

are called Remaining Performance Obligations (RPO) or Commitments. RPO are booming in an extraordinary way. RPO are revenues more than 1 year out (usually), but we do not know over how many years. Maybe the time-horizon has extended and that explains some of the increase. But still, the increase from mid-2025 is amazing.

Now we can wonder about the impact of Capex on RPO.

The chart below is based on 2 quarters average to smooth out quarterly variations. It calculates: the increase in RPO (2 quarter avg) divided by Capex. Dollar of new RPO divided by Capex-$.

This means that:

• $1 in Capex (1 million, 1 billion, as you wish) used to trigger $0.5 in committed future revenue (or ½ million, ½ billion, as you wish)

• Now, $1 in capex triggers $3 in committed future revenues.

Maybe that’s a naive illustration. Maybe there are “long and variable” lags between Capex and future revenues. Maybe the increase in RPO is linked to new product launches, not directly Capex. But still, committed future revenues are increasing a lot faster than Capex.

Valuations based on Consensus as off yesterday, which is ~20% too low. The stock is trading at 14.1x FY26 EPS, 7.1x FY27, ~5-6x FY28.

Supply tightness beyond FY27, incremental price upside (hikes and the mix) will have a lower impact on margins now (OPM already at ~81%), Capex could double in FY27 ($27bn to ~50) or ~40% of cash from operation.

Share

The table below shows Consensus as off yesterday. In yellow: FY26 is known since we have 4Q26 guidance. My FY27 assumptions are based on 4Q26 (50bn revenue plus 20% growth for FY27, 2H26 operating margin). There’s no creative assumption there. It reflects where Micron is in 4Q26. Consensus is 20% too low (at a minimum).

Share N. Baratte - TechStock01

From the conf call and prepared remarks

• data center revenue exceeded $25 billion in fiscal Q3, or an annualized run rate of over $100 billion

• DRAM and NAND industry demand continues to significantly exceed industry supply

• With respect to supply, our customers are recognizing that supply shortages in memory and storage will take considerable time to improve

• We have now signed 16 strategic customer agreements, or SCAs, across the data center, consumer and auto market segments. These SCAs provide customers with contracted supply assurance. The 16 signed agreements represent roughly 20% of our DRAM volume and a third of our NAND volume over this period. When completed, we expect approximately half or more of our company revenue to be under these SCAs.

• these agreements have a five-year term, from calendar 2026 through the end of calendar 2030. Automotive agreements generally have a three-year term.

• SCAs include four very large customers and three medium-sized customers. remaining agreements relate to smaller customers from the automotive industry

• The largest agreements generally have a ceiling price for existing products at the current CQ2 (calendar Q2) market price, and a floor price through the term of the agreement

• When all planned SCAs are executed, agreements with either fixed prices or price ceilings at or close to current CQ2 market prices are expected to be approximately 40% of our revenue

• 14 of the 16 SCAs that we have signed have a cumulative revenue at minimum price per our contracts of approximately $100 billion over the remaining agreement term

management stressed actual revenue will be “much higher”

• For our SCAs with price bands, the floor price enables a very robust gross margin for Micron, well above our peak quarterly margins in any past cycle.

• We expect tight conditions to persist beyond calendar 2027 as a result of AI-driven demand across all segments coupled with structural supply constraints.

• Even as we expect industry supply to improve gradually in 2028, we currently do not have line of sight as to when memory supply will be able to catch up with increasing demand

• Technology transitions are driving slower bit growth over time, wafer growth needs are significantly increasing cleanroom space and greenfield fab requirements, and HBM growth and increasing trade ratio with every new generation further pressures non-HBM supply.

• industry suppliers redirecting cleanroom space from NAND to DRAM and overall limited cleanroom space constrain NAND bit supply growth.

• Our 1γ (1-gamma) DRAM node and G9 NAND node are both ramping well and on track to become the highest volume nodes in Micron’s history. Development of our next-generation DRAM and NAND nodes are also progressing well and are on track to begin volume production in the second half of calendar 2027.

• We are leveraging our leadership DRAM and NAND nodes across our product portfolio. HBM4 12-high volume ramp is tracking twice as fast as HBM3E 12- high, and we have already shipped over $1 billion in HBM4 revenue. We expect to reach mature yields on HBM4 12-high significantly faster than HBM3E 12-high.

Management mentioned HBM TAM will crosse $100bn in ‘27 (pulled in from ‘28).

• Transitions like LP5 to LP6, DDR5 to DDR6 and newer generations of HBM all come with rising bit costs. • This trend, along with the ramp of significant greenfield capacity in the years ahead, is projected to cause the blended DRAM cost per bit to rise from current levels. Our customer SCAs provide for appropriate price premiums for such new products to be negotiated in the future.

• We now expect calendar 2026 industry server units to grow high-teens percent, above our prior expectations of low double digits, driven by mid-teens growth in traditional servers and even stronger growth in servers with AI accelerators. This increase in our server unit growth expectation is enabled by a modest reduction in average server DRAM content growth.

• In DRAM, we expect industry DRAM bit shipments in calendar 2026 to grow in the low- to mid-20s percentage range, slightly above our prior outlook.

• In NAND, we expect industry NAND bit shipments in calendar 2026 to grow approximately 20%, unchanged from prior expectations.

• We expect Micron DRAM supply to grow approximately in line with the industry supply growth, while Micron NAND supply grows somewhat less than the industry supply growth in calendar 2026.

• full-year fiscal 2026 capital spending to approximately $27 billion. We expect quarterly capex in fiscal 2027 to be above fiscal Q4 levels, with more than half the increase year over year in fiscal 2027 from construction capex as we pull in cleanroom capacity required to address long-term demand.

PAYWALL

Read more

Korea Semiconductor export (Memory) up 173% YoY. SKH up 250%, Samsung 180% YTD

Which are made of ~80% Memory and ~20% logic chips (Samsung LSI, foundry like DB HiTek, SK Key Foundry). Let’s say that the booming component is Memory, not the rest.

• April 2026: US$ 31.9 bn, up +173% YoY

• Year-to-Date (4 months): US$ 110.4bn, up +148% YoY

Nothing new… the boom continues obviously. Charts below show quarterly industry numbers, and operating margins. Triple bonus: ASP increase * mix improvement * margins increase.

TSMC: revenue up +30% YoY, stock up 48% YTD

• May 2026: NT$ 416bn (US$ 13.2 bn), up +30% YoY

• Year-to-Date (5 months): NT$ 1,961 bn (US$ 62bn), up +30% YoY

Mediatek: revenue slight decline YoY, stock up 200% YTD

• May 2026: NT$ 47bn (US$ 1.5 bn), up +5% YoY

• Year-to-Date (5 months): NT$ 243 bn (US$ 7.7bn), down -2% YoY

1Q26 smartphone sales (units) declined -3% to -6% YoY, depending on who you ask (IDC, Counterpoint). Quarterly smartphone units in million below:

Smartphone chips (SoC) declined by more than that: -8% YoY (Counterpoint), with “Recovery Expected Early 2028” (Counterpoint).

What they’re doing with the money?

It turns out that Google needs $80bn. What for? 1) AI Capex 2) pay tax. Seriously, to pay taxes.

“Alphabet intends to use the proceeds:

Capital expenditures to scale AI infrastructure and global compute

to facilitate an administrative change to how it meets tax obligations associated with vesting of employee equity awards. The company expects approximately $30 billion of proceeds will be used to meet these 2026 calendar year tax obligations”.

Demand was good, and the initial $80bn was upsized to $85bn.

Google did something quite unusual 2 days after the announcement (June 3^rd) which was to publish a justification. A full marketing spiel. It is here. It is called “Alphabet Investor Presentation June 2026”.

What do we learn here? For me personally, nothing much:

• Google’s PR team is using too much Gemini. A lot of pretty logos and illustrations – and little content.

• All the slogans that you’d expect:

  • “AI is the most profound platform shift of our lifetimes

  • unique opportunity… scaling our investments… significant growth opportunity ahead…

  • we’re innovating at scale with incredible velocity”

• Historical financials reiterated. No projections – of course no forecasts! that would be…. dangerous. But plainly zero forward looking nothing.

Should you have a look?

• Yes, if you want to grasp the extent of Google’s AI portfolio, research, products, etc = it is a good summary, it is impressive

• No, if you’re looking for new information

So why should you buy GOOG stock? Here is the spiel:

To be complete, a bunch of SEC filings here, here, here, here, and here.

Share

Space Exploration Technologies Corp, aka SpaceX

IPO prospectus “FORM S-1” here

“Use of proceeds:

1. expansion of our AI compute infrastructure

2. enhancements to our launch infrastructure and launch vehicles

3. increases in the scale and capacity of our satellite constellations”

Prospectus page 8-9: “expansion of our AI compute infrastructure” means building a few Fabs with Intel on earth, and Data Centers in space:

“AI Compute Infrastructure.

xAI has established a leading position in building and scaling terrestrial AI compute infrastructure… the first company to deploy a coherent gigawatt-scale AI training cluster.

The addition of Terafab, a chip manufacturing initiative with Tesla and Intel, aims to further extend our vertical integration to chip design and manufacturing to alleviate potential future chip shortages at SpaceX… the future of AI will be determined by the control of the physical stack

We Believe Orbital AI Can Accelerate Time to Power and Reduce Token Costs.

The Sun contains approximately offers the only truly scalable solution to the challenge of accelerating demand for compute relative to terrestrial energy constraints. The logical path forward is to move power-intensive AI workloads into orbit.

We Believe We Are Well-Positioned to Deliver Orbital AI Compute.

We believe orbital AI compute is an incredibly difficult technical challenge that only we can solve at scale in the near term. We are the only company that has already accomplished the key technical challenges associated with evolving connectivity satellites into AI compute satellites.”

Share N. Baratte - TechStock01

Tokens are too expensive

Microsoft’s AI Chief Says Anthropic Models Are Too Expensive

“The software company’s AI chief is working to build cheaper in-house models”

I’m sure it’s true but at the same time it’s propaganda: Google can do cheaper models than Anthropic. Show me.

Sam Altman says OpenAI’s top token spender uses 100 billion tokens a month — and they’re not even the world leader

Altman said the company was continuing to push its models and explore other ways to deliver “more value for less spend”.

He also said that the cost question came up quite suddenly. At the beginning of 2026, “the issue never came up”. Now, AI costs are “a huge issue”.

That’s just the latest 2 examples that I’ve seen, just this week. It is not a new debate but a growing debate - growing concerns.

Burden of proof

I’ll stop here. Because there is no concluding evidence for now – and maybe for the next 2 years: we cannot measure AI’s ROI. What we can see is this, using Google’s presentation:

1. revenue growth and margins accelerating

2. but Capex is exceeding Operating Cash flow, hence more debt / fund raising

3. how fast can unit cost decline?

Subscribe now

Previous reports

Hyperscalers: is AI Capex too high? What’s AI Return on Investment?

Hyperscalers: What is Consensus forecasting? Good or bad? how much Capex can they spend?

• Capex has increased like crazy, asset efficiency keeps declining (1$ of equipment is generating less revenues) – and has been declining for the past 10 years.

• headline AI Capex is not the entire story: margins are increasing, free cash flow is lower but solid.

• slightly negative FCF is tolerable for a couple of years. Debt levels are low. More off-balance sheet financing is possible

• on the other hand, I also do not think that the hyperscalers’ managements and shareholders have any appetite for prolonged negative Free Cash Flow, nor for high gearing levels.

Semiconductor market value. TSMC said $1.5tn by 2030. WSTS says $1.9tn in 2027.

• For the Semi industry to reach $2.5tn by 2030, how much AI Capex and revenues do we need by 2030?

• from 2026 to 2030, Hyperscaler’s revenues must more than double, operating profits must more than double to absorb that Capex and Semi value

What’s wrong with the outlook?

1. News says that Broadcom results were poor because ’26 AI chips revenues are guided at $56bn, but Consensus is at $57.6bn: 3% below! Or 3Q revenue guidance was also 3% below… I don’t think there’s a panic over this.

2. the firm maintained its ’27 AI revenue forecast “in excess of $100bn”, while pointing out in a convoluted way that if growth continues at 100% (2x)… you get $120bn.

“on the second half (of ’26), we keep the momentum going as we expect to see in 2027. What we will see in 2027 is continued growth of the level we’re talking about. And if you drive on that basis of what we’re seeing here, in the range of 2x what 2026 will be, I think you will easily see that 2027 will exceed very easily $100 billion in 2027”.

3. warning, again, that AI chip growth will dilute gross margins, and again in a convoluted way but it could be accretive to operating margin. You got to hear it to believe it:

“semiconductor business grows, just to reiterate, on a consolidated basis, relative to our software business, you’re going to have a decline in margins, right, a bit. You’ll have compression, but remember that it’s accretive because we have strong operating leverage, right? So our operating margins will stand up a bit over time to that”

Seriously, I am not sure that there’s bad news here. ’27 AI revenue between 100-120bn with a bias toward $120bn. Operating Margin won’t decline:

• in ’25 65.7%

• it’s up in ’26 so far at 66.9%

More important: TPU financing

CEO hinted that the TPU leasing business will become very large. That’s called the “AI XPV platform”. Details aren’t clear to me, i.e. little disclosure but refer to Bloomberg, another Bloomberg and Reuters. The idea is that Google agreed to “sell” TPU to Anthropic and OpenAI. Broadcom creates an “AI XPV platform” together with Apollo and Blackstone. These 2 provide the financing for the chips (i.e. they pay Google). Anthropic and OpenAI rent the chips, i.e. pay back Apollo and Blackstone – it’s a loan. And apparently Broadcom provides a “backstop”, i.e. guarantees the loans to Anthropic and OpenAI.

Ahhh another circular financing thingy like Nvidia with CoreWeave. You are financing your customers! Not exactly as Apollo and Blackstone are providing the loans, but there’s an unknown Broadcom guarantee in there.

The size looks very large:

“we are creating the AI XPV platform with Apollo and Blackstone and other leading investors to deploy more than 20 gigawatts of compute capacity through 2028. The first tranche of this platform valued at $35 billion is in fact currently being launched by Apollo.

Anthropic: we are providing access to Broadcom TPU-based compute of over 1 gigawatt. we entered into an agreement to access another 5 gigawatts of next generation TPU in 2027

OpenAI: production late 2026. contractual commitment to deploy 1.3 gigawatts in 2027, as part of the larger 10 gigawatts by 2029”

You can pick up from the conf call that ’27 AI chips revenue ($100bn) is equivalent to 10 Gigawatt, then.. 1GW = $10bn in AI chip revenue, and 20 GW means $200bn? That’s a minimum estimate!

I am hearing much bigger numbers. For example Nvidia GB200-based data center. 1GW = $38B total CapEx, of which:

• servers ~$23B

  • of which

  • ~70% is compute silicon or ~$16-18B in GB200

  • ~25% is networking

• Storage $10bn

• The building, power, air con, etc $5bn

My estimates for Amazon and Google’s Data Center is $33-35bn per GW

You can find lower numbers, looking at Nebius, Iren, CoreWeave but these firms aren’t clear about exactly what their Capex covers (ie land, building, only chips, etc). The range is $15-21bn / GW.

Conclusion: if Broadcom CEO is correct on “deploy more than 20 gigawatts of compute capacity through 2028”, the cost of that is at a minimum $200bn. Could be as high as $400bn.

Another possibility is simply:

Semi correction is starting – after a mad rally since early April. Amplified by news that are seen as bad news.

TSMC Warns Chip Supply Won’t Meet AI-Fueled Demand for Years. Yes, we know that.

Microsoft’s AI Chief Says Anthropic Models Are Too Expensive. Trashing your competitor?

From the conference call:

• semiconductors. Q2 revenue was a record $15 billion, up 79% year on year. Driving this growth was AI semiconductor revenue and a record $10.8 billion, up 143% year on year and above our outlook. Networking represented almost 40% of our Q2 AI revenue. Demand for XPUs and networking is simply insatiable

• During the quarter, bookings for AI semiconductors were over $30 billion, against the $10.8 billion we shipped.

• In the second half of 2026, we expect AI semiconductor revenue to double from the first half. Consistent with this trend in Q3, we expect AI semiconductor revenue to accelerate to $16 billion, up over 200% year on year.

• full-year 2026, AI semiconductor revenue $56 bn, up 180% YoY.

• fiscal year 2027 reiterate AI semiconductor revenue guidance in excess of $100 billion

  • on the same trajectory as we are seeing in the back half of ‘26.

  • you will easily see that 2027 will exceed very easily $100 billion in 2027, we are continuing to say that it will be over $100 billion in 2027.

• fiscal 2028 revenue growth to continue. We expect, in fact, ‘28 to be a substantial growth

  • for ‘27, we indicated about 10 gigawatts shipment in ‘27. That’s still very much intact. They will be shipping 10 -- we’re planning to ship 10 gigawatts in ‘27 and that nothing has changed, back-half loaded to that extent, yes, which really provides an interesting trajectory into ‘28 with this back-half trajectory. So ‘28, we expect a lot more gigawatts.

• we have been able to secure supply for our needs, ‘26, ‘27. Working on ‘28 and ‘29 right now

• Our visibility runs all the way to 2028 right now

• Google: long-term agreement to develop and supply multiple generations of TPUs and AI networking

• long-term agreement with Google: it’s a commitment that is very substantial in dollars, very, very substantial amount of dollars. Now, we also accept the fact that given the growth of AI compute by our partner Google, we fully expect that there’ll be some diversity of sources for them

• Anthropic: we are providing access to Broadcom TPU-based compute of over 1 gigawatt. we entered into an agreement to access another 5 gigawatts of next generation TPU in 2027

• OpenAI: production late 2026. contractual commitment to deploy 1.3 gigawatts in 2027, as part of the larger 10 gigawatts by 2029

• Meta: partnership for multiple generations of MITA XPUs. we expect to deploy 3 gigawatts through the end of 2028. Initial order for 1 gigawatt, includes XPUs and networking, will start in the second half of 2027.

• other two customers: expect shipments late 2026, accelerate in 2027. purchase orders for $6 billion.

• For scale up within racks, we enable direct attached copper based on an industry-leading 200G and 400GB SerDes, driving co-packaged copper with Ethernet and PC Express switches.

• For scale out between the racks, we have been shipping the industry’s only 100-terabit Ethernet switch, the Tomahawk 6, for over a year. We will now be taping out our next-generation 200-terabit switch this quarter.

• CPO, Co-Packaged Optics, 1.6-terabit DSPs, CW, and EML lasers, we are the de facto standard in the industry.

• Our strategic vision is to bring together Broadcom’s leading technology and investor partners with the strongest balance sheets to deliver at scale sufficient compute capacity at the lowest cost and power for the leading AI frontier labs, including Anthropic and OpenAI. To deliver this vision, we are creating the AI XPV platform with Apollo and Blackstone and other leading investors to deploy more than 20 gigawatts of compute capacity through 2028. The first tranche of this platform valued at $35 billion is being launched by Apollo.

• infrastructure software

  • Q2 revenue $7.2 billion up 9% YoY

  • Q3 revenue $8.9 billion up 31% YoY

• released VMware Cloud Foundation 9.1, focused on improving infrastructure efficiency, security, and support for enterprise, AI inferencing workloads. VCF 9.1 for on-prem cloud computing release heterogeneous compute support across GPUs and CPU architectures, including AMD, Intel, and NVIDIA platforms, enabling enterprise cloud customers to run AI Kubernetes and traditional virtualized workloads on a common private cloud environment.

• As the proportion of AI revenue significantly grows in Q3, we expect Q3 consolidated gross margin to be down to approximately 74%. This decline in gross margin does not represent a structural change in semiconductor margin, rather it reflects product mix between semiconductors and infrastructure software.

Revising up

Long long time ago, Mc Kinsey said that the Semi market will reach $1 trillion by 2030. Today, this forecast is irrelevant: it was published in 2022, the main driver was Automotive / Industrial, it was before AI, before HBM, before EUV explosive costs (i.e. TSMC 5nm wafer price was $17k. Tomorrow, 2nm price ~30k).

I only mention this because people still refer to it. For ex, TSMC in its May ’26 Tech Conference in HsinChu said: $1 trillion is too low, it’s $1.5 trillion. That’s the context of this slide below.

One month later… Memory contracts (HBM4, RDIMM / MRDIMM) for 2026 and 2027 are settled. Nvidia and AMD have reported and

• Nvidia CFO Colette Kress increased the “revenue visibility figure” from $500B over 2025–2026, to ~$1 trillion through 2027.

• Jensen went to Taipei – everybody is in Taipei, it’s Computex time and dropped $150bn:

  “Taiwan is incredible at technology manufacturing. This is the epicenter of the ecosystem... 4 years ago, 5 years ago, Nvidia was spending about 10, 15 billion dollars a year in Taiwan. Now we’re spending 100, going to 150 billion dollars in Taiwan each year”

• AMD Lisa Su went to Taiwan and AMD also has 2-3 years visibility:

  “we have very good visibility now into the deployments that are on track for 2027. And when I say good visibility, it’s visibility down to which data centers are the GPU going to be installed in”, “we’re now talking about ‘27 CPU demand, we’re talking about ‘28 CPU demand. And so that allows us to just plan much better as we go forward.”

  “If AI were a nine-inning baseball game, we’re probably only in the third inning... the global artificial intelligence boom remains in its early stages”, she said.

I am just saying that demand estimates are flying higher every day – what does it mean for total Semi industry value?

I wrote yesterday on TrendForce Memory value forecasts, going to $1.3 trillion in 2027.

I was going back to my spreadsheet on total Semi industry value but WSTS conveniently published their forecasts.

The numbers in blue cells are not published by WSTS, it’s me extrapolating based on WSTS’s charts.

Zooming in:

Sorry TSMC, it’s not $1.5tn by 2030, it’s $1.5tn in 2026. I guess we could guestimate $2.5tn by 2030.

In terms of revisions, and to visualize the steepness of estimates increase, WSTS (total Semi) and TrendForce (Memory only) are similarly ballistic. Here are the last 3 forecasts from WSTS:

Looking at the details of WSTS forecasts (table below):

• Similar to TrendForce, the biggest growth is Memory

• WSTS is a bit lower than TrendFroce but it’s in the ballpark. 2026 $804b versus 889b. 2027 $1.1tn versus 1.3tn.

• For Logic, growth is lower because the main driver is unit growth, and secondary driver is cost increase (cost because 3nm -> 2nm -> 1.6nm wafer price goes up ~20% by generation – but that not a price increase).

We can compare WSTS Logic and TSMC Consensus forecasts. It is “consistent”, with WSTS is lower and a bit too low, as it should be 1) TSMC will grow faster than many other Semi firms 2) TSMC analysts have a better grasp on capacity increase than WSTS on total demand.

So that’s fine. Again, it’s in the ballpark.

Let’s say it’s gonna happen. Let’s say we hit $ 2 trillion Semiconductor revenues in 2028 and $2.5 tn in 2030.

Subscribe now

Where is all that money going?

You’ve been told during the last earnings season, Agentic / Inference demand is going ballistic. I’m not going to copy tons of excerpts from transcripts. Just this from Microsoft, Satya Nadella:

“A new app platform is being born. You can think of agents as the new apps” “We are at the beginning of one of the most consequential platform shifts that will change the entire tech stack as agents proliferate and become the dominant workload.”

Remember the overall conclusions:

1. Inference demand is now bigger than Training demand — Broadcom was most explicit; Nvidia, AMD, and Intel all confirmed the shift. Agentic AI is driving a CPU renaissance — Intel and AMD both raised CPU TAMs.

2. “Agents are the new apps” Microsoft said it most cleanly. Amazon, Google, and Nvidia said it in different words. Every hyperscaler is building an agent deployment platform.

3. Inference cost reduction is the key unlock — Amazon, Google (TPU 8i), Nvidia (Vera Rubin at 35x better inference) and Microsoft (40% inference throughput improvement) are all racing on cost-per-token.

Important conclusion 1:

AI Capex is accelerating and the issue is not demand, not applications, it is the cost curve of token generation. It got to become 50x or 100x cheaper. My gut-feeling is that we are the the way to get there in a few years.

Nvidia’s 35x increase in Inference thruput (with Vera Rubin) means from one Gen to the next Gen of GPU.

Microsoft’s “40% improvement in inference throughput for our most used models across Copilot, driven by our software and hardware optimization work” means using the same existing stack.

Important conclusion 2: Hyperscaler’s revenues need to double, operating profits need to double by 2030

If we assume that 55-60% of Semi spending is on AI:

• $1.2 trillion in 2028, $1.5tr in 2030 for AI Semi

If we assume that Semis are ½ to 2/3^rd of Capex:

• $1.9 trillion in 2028, $2.3tn by 2030 for Capex

• 50% of total Capex is by hyperscalers =

  • $950bn Capex in 2028, Consensus is at 860bn

  • $1.1tr Capex in 2030

Hyperscalers want to maintain their cost structure – actually margins are increasing as Cloud revenues are increasing. But we should have some max burden for D&A/revenue (let’s say ~30%) and Capex/revenue (let’s say below 20%).

Then what does it mean for Hyperscalers revenues, profits – and by implication for the future AI application industry?

It means that from 2026 to 2030, Hyperscaler’s revenues must more than double, operating profits must more than double to absorb that Capex and Semi value.

This is not a FORECAST. I do not know if it will happen. What I am saying is: for the Semi market to reach $2,5tn in 2030, Hyperscaler’s revenues and operating profits must more than double.

Before you conclude that revenue / profits cannot double over the next 4 years, remember that Hyperscalers operating profits have doubled over the past 4 years.

If you’d had the courage to read all this, here is Conclusion 3:

• Semiconductors will outgrow Hyperscalers and AI SaaS for another ~2 ~3 years

• Within Semi, Memory outgrows Logic in 2026-27

• There will be an inflection point in ~2028 when Hyperscalers and AI SaaS will outgrow Semiconductors – big musical chair rotation

1Q26 in red

TrendForce has revised its Memory market value forecasts. 2026 forecast revised up 60%, 2027 revised up 50%

Previous forecast was published in Jan ’26. Total Memory value:

• 2026 US$ 552 bn or up 134% YoY

• 2027 $ 843bn or up 53% YoY

New forecast was published on 29 May ’26. Total Memory value:

• 2026 US$ 889 bn or up 278% YoY

• 2027 $ 1,283bn or up 44% YoY

2 things to say about this.

Why such a big revision? 3 main elements:

• Hyperscalers have revised up their Capex for 2026. Remember consensus was expecting Microsoft to spend $90-100bn, and Microsoft guided for $190bn. For example.

• Memory price increase is higher than expected. TrendForce was expecting 1Q26 average DRAM price to increase by 55% QoQ, but it was 90%. 2Q26 DRAM prices now expected to increase by 60% QoQ. This is mitigated somehow by HBM price decline with its cost curve / volume growth.

• Finally, unit demand is higher due to Inference / AI CPU growth. From TrendForce:

  “DRAM demand trends indicate that inference requests in Agentic AI systems are evolving from single queries into continuous iterative cycles. Furthermore, KV cache capacity is scaling proportionally with larger context windows. If recalculation is required, compute costs will increase exponentially. As a result, efficient KV cache management has become critical to AI inference performance, directly driving demand for HBM and DRAM.

  Agentic AI workloads are significantly increasing CPU requirements for scheduling, data preprocessing, and memory management. In next-generation AI server platforms, CPU-to-GPU ratios have gradually shifted from the traditional 1:8 toward 1:4 or higher. For example, NVIDIA’s NVL72 rack adopts a 1:2 configuration. This means rising CPU deployment is expected to expand server DRAM capacity requirements while simultaneously supporting procurement volumes and contract pricing.”

Which means that unit growth is not only HBM linked to GPU or Accelerator, but also MRDIMM (Multiplexed Rank DIMM) linked to the CPU. If you’re a purist, you can argue that some CPU will still use the conventional RDIMM (Registered DIMM). Ok, whatever: more CPU, more DDR capacity per CPU, higher speed, lower latency, etc which means that this is sucking up more wafer capacity than commodity DDR.

TrendForce Jan ’26 forecast

TrendForce May ’26 forecast

Share

You’re gonna say: OMG that’s big. Actually it is not!

The numbers above show a very large revision, but that’s a catch up for 2026 based on what I highlighted 1) higher capex 2) higher price increase 3) higher volumes.

Now, the problem is 2027.

I do not know what TrendForce forecasts by quarter – they don’t release that. But I have historical quarters (up to 1Q26, red bar below), I have 2Q and 3Q price increase estimates, so I can derive linear forecasts. Like this:

As you can see above, visually, growth is expected to collapse from currently ~250% YoY down to 0% by end of 2027. The revenue bars for 1Q-4Q27 are flattish.

This doesn’t make a lot of sense if we believe that:

1. Prices will remain flat, maybe won’t increase any longer, but certainly won’t decline. As off today, that’s what the whole industry is saying: memory shortages, 3-5 year contracts with fixed prices, pre-payments, new greenfield capacity won’t kick in before 2028, etc

2. Demand, capex, units deployed (AI servers, GPU, CPU) will continue to increase. At least there is an agreement that Agentic AI / Inference is at an inflection point and will accelerate. Refer to Nvidia, Google, Microsoft last conf calls.

So why on earth is TrendForce forecasting flat revenues?

Maybe because they don’t know what to forecast.

Maybe this forecast means: memory value cannot decline, the minimum forecasts are what you see in the charts above – but we don’t know how much upside.

Honestly, if I was doing that forecast for TrendForce, I’d be careful for a simple reason: memory price increase, both DRAM and NAND, have exceeded what everybody was expecting – including what the Memory makers were expecting 6-12 months ago.

Starting 3Q25, the magnitude of QoQ price increase was beyond one’s wildest imagination at 15-20% QoQ. Don’t even mention +90% in 1Q26 and +60% in 2Q26.

In the table below, which compiles ASP forecasts of TrendForce and Counterpoint, you can see something crazy hitting in 2-3Q26: server DDR (MRDIMM or RDIMM) will become more expensive than HBM ($ per GB). This isn’t “price speculation” or “shortages” alone, it’s also spec increase: speed increase, bandwidth increase, density increase.

Margins expansion: how much higher?

*** PAYWALL***

Subscribe now

Read more

Subscribe now

Bloomberg’s wonderful article – completely wrong: “Huawei’s semiconductor chief He Tingbo said the company will start making 1.4-nanometer chips by 2031 with its own “LogicFolding” technology”.

What Huawei did and didn’t say

What Mrs He said was:

• Huawei has developed a technique to design transistors that are closer together, that have better capacitance, better interconnect, less wiring, etc

• so that you can squeeze more transistors, design a higher transistor density – which is not the same as making smaller transistors

• so that the transistors can communicate faster

• and that will be equivalent to 1.4nm => really?

What Mrs. He did not say, but what she meant is: since Chinese firms cannot buy ASML EUV machines, Huawei needs to find a workaround, a new transistor design that doesn’t require EUV. And she did not say that Huawei will manufacture 1.4nm chips.

That was at an IEEE conference in Shanghai. Usually, the keynote speakers have a bit of something to show. Chip design? Chip performance? Some benchmarking? Here, Mrs. He had nothing to show. It doesn’t mean that Huawei has nothing – it can mean Huawei does not want to show anything.

The questions therefore are:

1) does Huawei have a new design methodology that the rest of the industry didn’t think about, this mean the IMEC consortium. IMEC roadmap below.

2) that can be manufactured without EUV lithography?

Huawei is a major R&D firm

Huawei published an annual report. It’s always a good read.

Huawei is a giant: US$ 126bn revenue in 2025

And a very big R&D organization: US$ 27bn in 2025. This compares to:

• ASML US$ 4.7bn

• TSMC $7.8bn

• Intel $14bn

• Nvidia $18.5bn

Which means: do not underestimate Huawei. Do not dismiss Huawei’s announcements as wishful thinking. Huawei had AI deep learning processors (Ascend 910) in 2019, slightly ahead of Nvidia.

But the EUV bottleneck has been, for now, insurmountable.

Huawei’s conclusion

“By 2031, the high-end chips HUAWEI designs based on the τ Scaling Law are expected to feature a transistor density that is equivalent to 14 Å (1.4 nm) processes”.

Ok, see you in 2031.

Why do we want smaller transistors?

A computing semiconductor (ie a CPU, a GPU) is made of small boxes called transistors. In each little box, you have a piece of information, a 0 or a 1. By combining the 0 and 1, you can do calculations, i.e. computing.

Why do we want smaller transistors?

If you can make smaller transistors, you can have more transistors in a chip (a CPU or GPU), which means:

1. More transistors = higher computing power

2. Denser transistors = faster computing speed

3. Higher transistor density = cheaper chips

Example: the principle of AI computing were established 70 years ago. The first GPU was launched 30 years ago. But it’s only in 2017 that TSMC could make small-enough, dense-enough, fast-enough transistors to enable model training (Tesla V100 - Volta Architecture and the first Tensor Cores). It’s only in 2022 that there was enough computing power for massive LLM training (Hopper and Transformer Engine).

This is critical. You could have a very good idea, for example AI computing. But your chips don’t have enough computing power. It’s not because you don’t know how to design the chip. You know. But TSMC can’t make it, because they don’t have the tools (ASML).

Moore’s Law is not a law

If you’re kind, you can say that “Moore’s Law” is an empirical observation. It’s a marketing slogan that says semiconductors get smaller and cheaper every 2 years.

Gravity is a law. If you drop an apple (on earth), it won’t go up. It will do down. It will be attracted by earth’s gravity.

Semiconductors do not get smaller and cheaper because it’s their destiny or because of a mysterious law of the universe. Advances in lithography make smaller semiconductors. Plus many other things, but let’s say the critical enabler of smaller transistors is lithography.

Hence the US restrictions on China buying advanced ASML lithography machines called EUV.

SMIC can make 7nm, maybe 5nm but not smaller nodes

The main Chinese manufacturer of chips, SMIC, can buy immersion lithography machines that can make transistors down to 5nm using a manufacturing technique called multi-patterning. It can be done but multi-patterning costs a lot and has poor yields. A better solution is using EUV machines: you’ll have a lower cost and higher yields – but if you can’t buy EUV, you still can make 5nm chips. That’s what SMIC has been doing since 2020.

EUV is the glass celling, for now.

Left: SOXX versus IGV, indexed at 1 on 10 Oct 2025, up to 21 May ’26. I picked 10 Oct 25 because that’s 1) roughly the previous peak 2) a month before IGV started to decline.

Right: same but indexed at 27 Mar ’26, the recent bottom.

Same in percentage:

Semiconductors – the SOXX

I wrote about this here last week.

Main conclusions:

• The SOX is up ~60% since end of March ’26, on the back of massive earnings. Valuations are at average – not high.

• Memory sector: highest earnings growth, highest stock performance, and yet the lowest valuations. The cyclical vs structural argument won’t be settled anytime soon, but can offer a lot of upside when the market gets convinced by the structural argument.

• Since it’s complicated to buy Korean stocks, I just heard of a Memory ETF named DRAM that has ~50% exposure to SK Hynix and Samsung.

• Within the big caps, NVDA and TSMC are now abnormally cheap.

• Most expensive: ARM and Intel.

This is the SOXX ETF below. On the left, its share price – which we all know. On the right, its market cap which is the same profile as the share price. That gives an idea ofg the massive fund inflow into Semiconductor over the past 2 months. End of March, market cap was $20bn. Yesterday $35bn.

Hyperscalers: is it time to buy Microsoft?

Subscribe now

Below the Paywall

The case for Microsoft

Software / SaaS firms: what’s interesting? 4 stocks that look good

Read more

Earnings upward revisions will keep coming:

• Forecasts: Consensus is too low. My forecast show 20% upside to Consensus for FY28 and 40% for FY28.

• Valuations: Nvidia is trading below average PEx, based on a Consensus that’s ~20-40% too low. On my forecasts, the stock is trading at 16x FY28 and 11x FY29 EPS

Is there nothing to worry about? We always worry about everything.

For Nvidia, let’s boil it down to 3 points:

1. Is Nvidia losing market share? Yes, we know that since 2 years ago. Google TPU is growing faster than Nvidia GPU. AMD and Broadcom are growing faster than Nvidia GPU.

2. Will AI spending shift to CPU for Inference? Yes, Inference CPU will grow faster than Inference GPU.

3. Is AI making money… just wait for OpenAI and Anthropic IPO filings. Soon.

Nvidia understands the concerns, so Jensen says:

“NVIDIA is uniquely positioned at the center of (the AI) transformation as the only platform that runs in every cloud, powers every frontier and open-source model… from hyperscale data centers to the edge”.

This means: I know that I’m losing market share, but I’m at the center of all the infrastructure.

“The buildout of AI factories — the largest infrastructure expansion in human history — is accelerating at extraordinary speed”.

This means: I know that I’m losing market share, but the pie is growing a lot faster than you know.

“Agentic AI has arrived, doing productive work, generating real value and scaling rapidly across companies and industries”.

This means: Agentic will generate revenues, don’t worry.

NVIDIA is changing its reporting format.

We will now have:

• Data Center

  • Hyperscale

  • ACIE: AI Clouds, Industrial and Enterprise

• Edge Computing including “all the rest”: PCs, game consoles, workstations, AI-RAN base stations, robotics and automotive.

That’s a good thing. People keep asking the Data Center breakdown between hyperscalers and the rest. That’s more interesting than the previous breakdown, compute and network.

Little gift to thrift investors

• A new share buy back: Board of Directors approved an additional $80.0 billion to the Company’s share repurchase authorization, without expiration.

• A nice dividend: NVIDIA is increasing its quarterly cash dividend from $0.01 per share to $0.25 per share of common stock. That’s a 13% pay-out ratio on Non-GAAP EPS.

And yet. Nvidia generated $50bn of operating cash in 1Q27. Nvidia has $80bn of cash and equivalent in April ‘26. They don’t know what to do with the money – and returning it.

Share

Leave a comment

Below the Paywall

• 1Q27 results

• 2Q27 guidance

• Consensus is too low. My forecast show 20% upside to Consensus for FY28 and 40% for FY28.

• Valuations. On my forecasts, the stock is trading at 16x FY28 and 11x FY29 EPS

Reminder:

• FY28 ends in Jan-28 so that’s basically 2027

• FY29 ends in Jan-29 so that’s basically 2028

Read more

The SOX is up ~55% since end of March ’26, on the back of massive earnings

The SOX is up ~55% since March-end. Many stocks have doubled. The 1Q26 earnings season was massive; forecasts have been revised up. For the 55 stocks in my SOX Index, net income growth was 72% in 2025, forecast at 148% in 2026, another 44% in 2027. We haven’t seen that level of earnings growth since 2010, but that was a recovery after a nasty crash.

The SOX is not expensive. Charts below show my SOX Index incl. 55 stocks.

On trailing earnings (this means net income of the past 4 quarters, or 2Q25-1Q26), valuations were at average end of March and now at +1 standard deviation.

On forward earnings (this means the net income of the next 4 quarters, or 2Q26-1Q27), valuations were low in March at -1 standard deviation, and still below average today.

Share

How much net income growth?

Chart and table below show the sum of net income growth for the 55 stocks in my SOX Index.

• 148% growth in 2026. We have little uncertainty over 2026, baring a global recession (ie oil, Iran, inflation, rates), but probably little upside as well. The “full capacity” or “shortage” message from the manufacturers of chips (TSMC) and Memory (MU, SKH) has been ingested. A recovery in automotive and industrials doesn’t really move the needle, it’s too small. A -15% decline in PC and Smartphone volumes (Memory shortages or unaffordable pricing) should be factored in. The question is: can production decline by more than -15%.

  • Omdia on smartphone -15%,

  • Counterpoint smartphone -12%

  • IDC on computers -11%

  • Gartner on computers -8%

• 44% growth in 2027. Comprehensibly, Consensus expects a slowdown in 2027, and I have no idea about the magnitude of the slowdown. It will start from hyperscalers capex outlook and we won’t have any indication about this for another 6-9 months.

• 13% growth in 2028. Consensus is doing its normal thing: if you’re an analyst you have no incentive to forecast big growth. You want to increase your forecasts, you don’t want to revise down, when you have Buy ratings on stocks.

If we accept the premise that Semi growth is driven by AI demand (again, the rest is becoming too small to move the needle), there is an obvious disconnect between low Consensus forecasts and Jensen Huang saying that Inference / Agentic demand will be a lot larger than Ai capex so far. You can read this from TMT Breakout:

TMT Breakout

TMTB: Nvidia's Jensen (NVDA) and Michael Dell (DELL) Key quotes

Jensen and Michael were interviewed together today at Dell World in Las Vegas. Some interesting quotes, especially with NVDA EPS coming up on Wednesday…

Read more

4 months ago · 21 likes · TMT Breakout

By segment, no surprise:

• Memory is driving the bulk of growth in ‘26 and ’27.

• Followed by chip vendors ie AMD, Broadcom, Nvidia.

Note on composition:

• iShare SOXX ETF comprises 30 stocks

• my own SOX index comprises 55 stocks

• Semi Production Equipment (SPE) segment: 11 stocks, AMAT, ASML, etc

• Manufacturing segment: 8 stocks, GFS, TSMC, etc

• Chip vendors segment: 22 stocks, both IDM and fabless, AMD, Intel, NVDA, TXN

• Memory segment: 6 stocks, MU, SK Hynix, SNDK, etc

• Theres another segment of 8 “components” stocks that I do not show here: wafers (Sumco) and EDA (Cadence).

Subscribe now

The biggest money spinners

The table below shows the top-20 largest net income, sorted based on 2028.

wide divergence in performance and valuations

stock price appreciation since end March ’26, by segment:

PAYWALL ***

Subscribe now

Read more

The Semiconductor Production Equipment (SPE) sector is made of 2 different segments:

Wafer-Front End WFE equipment makers. They make equipment to manufacture the wafers inside the Intel, SK Hynix, TSMC fabs : AMAT, ASML, KLA, LAM, Tokyo Electron.

Back-End equipment makers, or Test & Packaging T&P. The equipment are used to test (Advantest, Teradyne) and package the semiconductors (ASMI, ASML, BESI, Hanmi, K&S). The equipment are used by Intel (EMIB), TSMC (CoWoS) and also 3^rd party back-end forms: Amkor, ASE, etc.

This report is about the 5 big Wafer Front-End equipment firms: AMAT, ASML, KLA, LAM, Tokyo Electron.

Everybody is revising up

We have a higher industry authority in the SPE sector, so let’s look at its forecasts first.

SEMI.ORG WFE forecasts: modest, boring and wrong

2026: 126bn up 9% YoY

2027: 151bn, up 20% YoY

The latest press releases are here and here.

These numbers are too low, and the 5 big WFE firms have revised up their forecasts and provided higher estimates. The only thing interesting in Semi’s estimates is that 2027 growth is much higher, 2x higher than 2026. Could be true – it’s directionally correct.

This is semi.org data:

Share

Revenue / market value estimates of the 5 big Wafer Front-End equipment firms

AMAT

Didn’t cite a total WFE dollar figure, but said their own semiconductor equipment business will “grow more than 30% this calendar year.”

2027–28: “I’m having constant conversations with customers, and they’re looking at ‘27 and they’re looking at ‘28.” Described leading-edge foundry-logic, DRAM, and advanced packaging as accounting for “more than 80% of the year-on-year growth in total WFE spending in 2026 and a similar profile in 2027”, “sustained, multi-year revenue and profit growth.”

ASML

No WFE dollar figure cited.

Has increased its 2026 revenue range from EUR 30-40 bn (mid-point 35bn) to EUR 36-40 bn (38bn)

Shipping “at least 60” low-NA EUV systems in 2026, growing to “at least 80” in 2027.

KLA

The 2030 revenue target is based on ~15% growth Cagr. This is based on WFE market at $215B ±$20B.

WFE 2025 baseline ~$120B.

2026 now “expected to exceed $140 billion” — increased from the $135–140B

2027: “today we expect the 2027 year-over-year growth rate to be higher than 2026.” CFO Bren Higgins: “there’s no question ‘27 is going to be a massive buildup.”

LAM

2026 WFE $135bn (January guidance), now increased to “$140 billion with a bias to the upside.”

From $116bn in 2025, $140bn means up 20% YoY

2027: “WFE is going to be nicely growing next year”, the firm declined to give a specific number but “firmly” positive.

Lam flagged ~$40B in NAND conversion spending needed over the next few years, with the majority now pulled forward to before end of 2027.

Tokyo Electron

Tokyo Electron is the most bullish with WFE for both CY2026 and CY2027 in the range of “$150–170 billion per year”. CEO Kawai on the $170B upper end: “I think $170 billion level of WFE market is achievable when I look at current customers’ investment plan.” Let’s assume this means:

2026 WFE $150bn, that’s 30% YoY

2027 WFE $170bn, that’s 13% YoY

The “range” is 20%+ growth versus 2025, that must be 2-year Cagr

My conclusion is:

We take all this, we remember ASML and KLA 2030 estimates, we add the semiconductor market value (I’ll get back to this in a few days), and we have a range of WFE spending i.e. revenues of AMAT, ASML, KLA, LAM, Tokyo Electron, plus a few smaller firms, that looks like this:

• You could say that it’s great, so much growth

• You could also say that peak growth in 2027 is not good for stock prices – especially after the huge rally of the past month / year

• But if you think this, you also have to remember that 2030 is very far away and nobody has an incentive to provide targets that they can’t reach.

In summary, there are 2 big positives and 2 negatives to watch carefully

Positive 1: the end-demand estimates keep increasing. That means estimates of the size of the semiconductor market keep going up. Yesterday, TSMC said $1.5 trillion by 2030.

You can read Jeff’s summary here, here.

This means something like this, my estimates to get to $1.6 trillion by 2030.

PAYWALL BELOW

Subscribe now

Read more