AI Chip Peak Check: Demand Still Rising, Risk Building in 2027-2029 Supply and Funding
DRAM and NAND contract prices are still climbing and HBM suppliers hold the edge. Power shortages, AI financing costs and simultaneous memory expansion are the risks for later years.
Key changes for Oct. 5-10, 2026. Separates the physical cycle, share prices and cash flow. Written for long-term ETF investors. Based on public information through Oct. 10.
Bottom line: The physical cycle is still rising. DRAM and NAND contract prices keep climbing, and HBM suppliers keep the upper hand. But AI financing costs, power shortages and simultaneous memory expansion are raising the risk for 2027-2029. There is not enough evidence to call an overall peak or rollover.
Overall read: physical demand rising; valuations stretched; mid-term oversupply under watch.
Key numbers
- +54.6%: TSMC September revenue, year over year (reported Oct. 8)
- 107.4 trillion won: Samsung Electronics third-quarter preliminary operating profit (Oct. 8)
- +10% to 15%: TrendForce forecast for fourth-quarter conventional DRAM contract prices (Sept. 30)
- 32 GW: Morgan Stanley estimate of the net US data center power shortfall through 2028
01. This week's key changes
- Earnings strength confirmed. TSMC's September revenue was NT$511.86 billion, up 54.6% from a year earlier. Samsung's third-quarter preliminary operating profit was 107.4 trillion won, up 782.5%. Both point to strong demand for leading-edge logic and memory.
- Advanced packaging expansion firms up. GlobalFoundries and TSMC signed a $2 billion contract (Oct. 8) for US-made silicon interposers used in CoWoS packaging. Volume is due to ramp in the first half of 2028. The deal does not add supply now.
- Power constraints threaten order timing. Morgan Stanley estimates a net US data center power shortfall of 32 GW through 2028, or 34% of demand. Downstream suppliers such as memory and optical makers face delivery adjustments if sites switch on late.
- AI credit market caution rises. The Financial Times reported Oct. 10 that AI-related borrowing fell to $23 billion in September. Investors are scrutinizing long-term returns and debt loads. That should not be read as falling AI demand.
- Next-generation EUV roadmap. ASML and Zeiss say Hyper-NA could be ready in about 10 years. It is a long-term option, not a near-term substitute for High-NA demand.
02. Scorecard
| Cycle | Call | Evidence and counter-signals |
|---|---|---|
| Nvidia, AMD GPUs | Rising | Nvidia's latest quarterly revenue up 106%, gross margin 75%. Customers' in-house chips are a long-term competitive pressure. |
| Broadcom, custom ASICs | Accelerating | Broadcom AI semiconductor revenue up 221% year over year. Demand is diversifying across TPUs, Trainium and others. |
| HBM4, HBM3E | Supplier advantage | Customer qualification and richer mix. A review of 8-high stacks is a new warning. |
| Server DRAM | Strong | RDIMM shortage; HBM taking a larger share of wafers. |
| PC, mobile DRAM | Prices strong, demand weak | Prices still rising, but consumer demand and memory content per device are under pressure. |
| Enterprise SSD | Strong | 2026 bit demand forecast to grow more than 80%; fourth-quarter price gains expected to widen. |
| Client SSD, consumer NAND | Demand fragile | Purchases driven by inventory drawdown, lower PC SSD capacities. Contract prices still rising. |
| TSMC, CoWoS, 2nm | Strong | September revenue up 54.6%, 2nm ramping. New CoWoS capacity arrives after 2028. |
| ASML, AMAT, LRCX, KLAC, TEL | Expanding | ASML installed-base service revenue rose. Little new order disclosure from the others this week. |
| Hyperscaler AI capex | Expanding | Cloud revenue and backlog rising. Capex growth and ROI are separate questions. |
| Free cash flow, depreciation, credit | Caution | Microsoft cloud gross margin fell, Oracle free cash flow is negative, AI borrowing cooled. |
| Power, cooling, optical | Bottleneck persists | Power shortage is both a delay risk and an infrastructure order opportunity. |
| CXMT, YMTC | Mid-term supply risk | Growing in conventional DRAM and NAND. Distinct from a direct threat to top-end HBM4. |
| Share prices, valuation | Expectations high | Memory stocks pulled back despite record profit. A cycle peak and a stock correction are different events. |
03. Memory pricing and HBM
Price direction: keep the pieces apart
- Conventional DRAM: Fourth-quarter 2026 contract prices are forecast up 10% to 15% quarter over quarter. Prices keep rising at a slower pace. That is not a turn lower.
- Total NAND: Fourth-quarter contract prices are forecast up 15% to 20%. Strong enterprise SSD coexists with weak client SSD demand.
- HBM ASP: TrendForce forecasts a 2027 blended ASP up 121% year over year. The forecast includes a higher share of pricey HBM4 and HBM4E. It is not the same as a confirmed 121% rise for like-for-like products.
HBM contracts and qualification
- HBM4 qualification: Reports say Samsung, SK hynix and Micron are all qualified for Nvidia's Vera Rubin. Actual shipments, yields and customer share by supplier need separate confirmation.
- Competitive picture: Second-quarter 2026 HBM revenue share was SK hynix 50%, Samsung 33%, Micron 18% (rounding totals 101%). Samsung is closing in; SK hynix still leads.
- Long-term agreements, take-or-pay: Micron's long-term customer commitments and deposits have grown. Price floors and take-or-pay terms by customer cannot be confirmed from public data. Remaining performance obligations or deposits should not be read as price-guarantee contracts.
- New warning, HBM downspec: GPU and ASIC makers are considering 8-high stacks instead of 12-high. This reflects system cost and supply limits. It is not a demand collapse, but the possibility of fewer HBM bits per GPU needs separate monitoring.
Rule of thumb: Contract prices can rise while spot prices fall. Strong total NAND contract prices do not prove a client SSD order recovery. A higher blended HBM ASP does not prove higher prices for comparable products.
04. Capex and the 2027-2029 oversupply risk
Now: investment is needed. Micron's fiscal 2026 net capex is $27.37 billion, up about 98% from $13.80 billion in fiscal 2025. The spending is tied to real customer demand and a richer product mix.
Later: a supply shock is possible. If leading-edge DRAM expansion at Samsung, SK hynix and Micron overlaps with higher CXMT output of conventional DRAM, pricing power could weaken in 2028-2029. Completion plans and actual operation and yields are different things.
- Structural constraint: HBM's share of DRAM wafer capacity is forecast to rise from about 20% in 2026 to about 30% in 2027. That could squeeze conventional DRAM supply.
- Order of confirmation: capex announcement, equipment move-in, wafer starts, stable yields, customer qualification, actual bit supply, then ASP impact.
- Warning combination: a surge in new bit supply, rising customer inventory, and contract price growth that drops sharply or turns negative.
05. EUV, equipment, foundry, CoWoS
- ASML: Second-quarter sales were 9.33 billion euros. Installed Base Management was 2.762 billion euros, up about 11% quarter over quarter. It plans to raise low-NA EUV capacity 30% in 2027. Third-quarter results are due Oct. 14. New orders and China DUV share get reassessed then.
- High-NA: Reports say Intel adopts first, Samsung and SK hynix in 2028, TSMC in 2030. Hyper-NA is a roadmap of around 10 years.
- TSMC: 2nm was 3% of second-quarter revenue, and overall gross margin was 67.7%. September revenue rose 54.6%. Third-quarter gross margin, 2nm yield and CoWoS utilization are due with results on Oct. 15.
- CoWoS: GlobalFoundries' $2 billion interposer deal signals a wider US advanced packaging chain. Volume ramps in the first half of 2028, so it should not be counted as current supply relief.
- Samsung, Intel Foundry: Technology progress should be judged separately from outside customer wins, yields and utilization. Improved profit at Samsung's foundry is still unverified.
- Applied Materials, Lam, KLA, Tokyo Electron: Likely structural beneficiaries of leading-edge, etch, deposition and inspection demand. There is no confirmed, company-level order disclosure this week to show all of them strengthening together. The next earnings reports should show the memory customer share of investment.
06. GPUs, ASICs, GPU cloud
- Nvidia: Fiscal 2027 second-quarter revenue was $96.22 billion (up 106%), data center $89 billion (up 117%), gross margin 75%. Vera Rubin production is expanding. The key is whether networking and optical revenue and GPU margins hold together.
- AMD: GPU alternative and rack-scale platform expansion. Cooperation with Samsung on foundry and memory is reported as under review, which is different from a signed contract.
- Broadcom, in-house chips: AI semiconductor revenue was $16.7 billion (up 221%), with $21.7 billion guided for next quarter. Google TPUs and AWS Trainium compete with Nvidia's share but add demand for TSMC, HBM and packaging.
- GPU cloud pricing and utilization: Differences by provider, generation and contract length are large. Public list prices alone cannot confirm rising rents or utilization across the market. Watch actual long-term contract prices, renewal rates, utilization, and canceled or delayed volume.
07. Hyperscaler AI capex and ROI
Microsoft (revenue evidence). Fiscal fourth-quarter Microsoft Cloud revenue was $59.3 billion (up 27%), capex $41.0 billion, free cash flow $19.6 billion. Cloud growth and cash generation are both confirmed.
Oracle (capital intensity risk). Fiscal 2027 first-quarter IaaS revenue was $7.4 billion (up 121%) and remaining performance obligations $664 billion. Capex in the same quarter was $28.5 billion and free cash flow was negative $5.4 billion. A contract backlog is not profit or cash.
- Depreciation pressure: Microsoft Cloud gross margin was 65%, versus 68% a year earlier, reflecting AI infrastructure spending and product mix. Company-wide depreciation and amortization was $11.0 billion versus $9.3 billion.
- Alphabet, Amazon, Meta: Conversion of AI capex into revenue varies across cloud, advertising and AI products. The test is not capex growth but whether incremental gross profit, free cash flow, depreciation, leases and debt improve together.
- Neoclouds (CoreWeave and others): Maturity mismatch risk between long GPU contracts and financing leases or secured loans. Contracted usage should be separated from actual GPU utilization.
- Credit markets: AI-related borrowing reportedly fell to $23 billion in September. Investors are demanding a bigger premium for debt, collateral and ROI risk.
ROI call: Rising revenue and backlog alone cannot show AI investment has paid off. GPU depreciation, power contracts, interest and lease costs need to be compared with actual paid usage.
08. Data center power, cooling, networking
- Power: Morgan Stanley's 32 GW shortfall through 2028 is a forecast model, not confirmed outages or canceled contracts. Grid connection delays could push back equipment deliveries and the timing of cloud revenue recognition.
- Beneficiaries: Transformers, switchgear, liquid cooling, power management, onsite generation, gas power and fuel cells. Backlog conversion to revenue and project permitting still need to be checked.
- Optical: Copper cable losses grow as links move to 1.6 Tbps. TrendForce (Oct. 8) expects wider competition between VCSEL-based NPO and CPO technology. That is an opportunity for optical module, laser and fiber makers, though not every roadmap has turned into revenue.
- Metrics: the gap between secured MW/GW and energized MW/GW, grid connection delays, cooling installation rates, GPU rack utilization and paid usage.
09. Chinese memory competition
- CXMT: Estimated at about 10% of global DRAM revenue in the second quarter of 2026. Customer wins and productivity gains in conventional DRAM and LPDDR are a pricing risk. They should not be equated with global qualification and yields for top-spec HBM4.
- YMTC: NAND expansion and more domestic customers are a mid-term variable for conventional and high-capacity NAND supply. Research firms use different bases, so no single current global share figure is given this week.
- Keys for 2027-2029: actual bit output from new fabs, yields, absorption inside China, global export prices, and changes in US and Chinese regulation.
10. Wall Street view and the valuation debate
Bull case: physical shortage and earnings. Record sales and profit at Samsung, Micron and TSMC, pricing power in server DRAM, HBM and enterprise SSD, and growth in GPUs and ASICs together.
Bear case: expectations, financing, supply. Shares fell despite Samsung's record preliminary profit. The core debate is whether peak earnings last, 2028 capacity additions, and rising AI credit costs.
A share price pullback is not enough to show demand has peaked. Strong earnings do not guarantee further gains for richly valued stocks. There was not enough data this week to check forward P/E and PEG consistently across the market, so no firm "bubble" call is made.
11. What long-term ETF holders can do
| Action | Assets | Criteria |
|---|---|---|
| Core hold | US semiconductor ETFs (SOXX, SMH), broad Nasdaq ETFs | Keep long-term core exposure until HBM, GPU, ASIC and foundry demand slow together |
| Complementary hold | Korean semiconductors with HBM exposure, AI power, cooling and networking | Confirm revenue and free cash flow conversion, not just backlog; check for overlapping holdings |
| Restrain new buying | Surging memory, equipment and optical themes | Stick to set weights and staged buying rather than chasing after strong earnings |
| Manage weights | Leveraged semiconductor funds, high-debt neoclouds, single-theme NAND | Satellite holdings, not core; beware daily reset and volatility decay in leveraged products |
| Reduce on conditions | Overconcentrated semiconductor exposure | Reassess in stages if three or more of HBM ASP, DRAM contract prices, TSMC gross margin, ASML orders and hyperscaler free cash flow worsen at once |
2028: a checkpoint, not an automatic sell date
- Good case: New wafer supply is absorbed by AI inference, ASIC and server DRAM demand, premium HBM ASPs hold, CoWoS utilization is stable and cloud free cash flow recovers.
- Bad case: The three memory makers and China raise output together, customer inventory builds, HBM downspec spreads, GPU cloud rents and utilization fall, and depreciation and interest costs jump.
- Action: Adjust ETF weights on measured values of these conditions, not on the date itself.
12. Next week's checklist
The biggest inflection points next week are equipment and foundry earnings.
- Oct. 14, ASML third quarter: new EUV and DUV orders, installed base revenue, China share of sales, 2027 capacity plans, any order cancellations or delays.
- Oct. 15, TSMC third quarter: leading-edge gross margin, 2nm yield and customer demand, CoWoS investment, 2027 capex and order visibility.
- DRAM and NAND prices: upward or downward revisions to the fourth-quarter forecasts (10% to 15% and 15% to 20%), and gaps between spot and contract prices.
- HBM4 pricing and specs: wider use of 8-high versus 12-high, HBM4E pricing, and public evidence of customer qualification and take-or-pay terms.
- Hyperscalers and neoclouds: AI credit spreads, large financings, GPU rents, renewal rates and utilization, power connection delays.
- Oct. 29, Samsung final results: actual HBM4 revenue and yields, foundry losses, mobile cost burden, 2027 memory capex.
13. Conclusion
Physical cycle: still rising. DRAM and NAND contract prices are up, premium HBM supply is tight, TSMC revenue is strong and GPU and ASIC demand is growing, all confirmed together.
Warnings: clearer than before. PC and smartphone memory strain, the HBM downspec review, a surge in memory capex, US power shortages, a cooling AI debt market and cloud depreciation pressure.
ETF response: hold the core, restrain chasing, check overlapping exposure. Do not fix 2028 as a sell date. Treat combined deterioration in HBM ASP, DRAM contract prices, TSMC gross margin, new ASML orders and hyperscaler free cash flow as the physical-peak alarm.
Insight Times Editorial Desk
