Claims ledger: hardware and compute market#
Status: Reference Last verified: 2026-09-26 Canonical for: every external claim our price projections rely on, with our verdict
This is the analysis layer. Each row is a claim from a source note, judged on its mechanism and on whether parties with different incentives confirm it. Our projections (buy-now analysis) cite claim IDs, so when a claim turns out wrong we can find every conclusion that depends on it.
Verdicts:
- holds: confirmed by independent parties or by the mechanism;
- doesn't hold: contradicted by better evidence;
- pending: plausible but unconfirmed.
Kinds: observed, forecast, or opinion.
Memory: DRAM#
| ID | Claim | Kind | Corroborated by | Verdict | Would be proven wrong by |
|---|---|---|---|---|---|
| SA-memmania-1 | DRAM density grew only about 2× in a decade (against about 100× at peak) | Observed | Industry roadmaps; fabs now add bits through wafers, not shrinks | Holds | n/a (historical) |
| SA-memmania-3 | AI takes about 70% of DRAM wafers by 2027 (12% in 2023) | Forecast | TF-0730-1; Micron guiding tight "well beyond 2027" | Pending (read second-hand) | Hyperscaler capex cuts; HBM orders cut in the 2027 negotiation |
| SA-memmania-4 | DDR margins reached HBM contract levels | Observed | Retail and contract price rises (SA-rental-3) | Pending (wording ambiguous) | 2027 HBM contract prices set well above DDR |
| SA-silicon-5 | HBM uses about 3× the wafer of commodity DRAM; HBM4 and HBM4E worsen it | Observed | TF-0730 (HBM wafer intensity) | Holds | n/a |
| SA-rental-3 | DDR5 contract about 5× and LPDDR5 about 4× year on year (Q1 2026) | Observed | US retail 32 GB kits $80 → $425–560; Gartner's +130% DRAM plus SSD | Holds | n/a |
| TF-0730-1 | DRAM undersupplied about 1–2% in 2026, widening in 2027 | Forecast | SA-dwarkesh-1; Micron | Holds | DRAM spot falling below contract; Micron gross margin falling sharply |
| TF-0730-2 | New DRAM capacity 2H 2027 at the earliest; substantial output in 2028 | Forecast | SA-dwarkesh-1; fab build times (about 2 years) | Holds | Announced fab slips (later) or pull-ins (sooner) |
| TF-0730-3 | DRAM prices elevated through 2027 | Forecast | SA-dwarkesh-1 (rising through 2027–28) | Holds | An AI capex correction in 2027 |
| SA-dwarkesh-1 | Memory prices keep rising through 2027–28 | Forecast | TF-0730-2, TF-0730-3 | Holds for 2027; pending for 2028 | 2028 fab output landing on time plus softer AI demand |
| SA-cxmt-1 | CXMT at about 350k wafers/month (2026) → 420k (2027) → 500k (2028) | Forecast | Other reports of about 350k by end 2026 | Pending | CXMT capacity or equipment-control news |
| SA-cxmt-2 | CXMT bit share about 9% → about 12% by 2027 | Forecast | Consistent with SA-cxmt-1 at a lower bit density | Holds (as the planning figure) | n/a |
| SA-cxmt-3 | CXMT cost per bit 30%+ above incumbents | Observed | Older process (1z-equivalent) | Holds | A CXMT node advance |
Memory: NAND#
| ID | Claim | Kind | Corroborated by | Verdict | Would be proven wrong by |
|---|---|---|---|---|---|
| TF-0730-4 | NAND supply turns positive in 2027; prices under downward pressure | Forecast | Layer-count scaling is still cheap, unlike DRAM | Holds | Enterprise SSD demand surge (AI inference storage) |
Logic and GPUs#
| ID | Claim | Kind | Corroborated by | Verdict | Would be proven wrong by |
|---|---|---|---|---|---|
| SA-silicon-1 | TSMC N3 is the binding logic constraint | Observed | SA-dwarkesh-2; TSMC sold-out reports | Holds | TSMC N3 capacity additions or AI order cuts |
| SA-silicon-2 | AI takes about 86% of N3 in 2027 | Forecast | SA-dwarkesh-2 (Nvidia 70%+ of N3) | Holds | As SA-silicon-1 |
| SA-silicon-4 | Nvidia moves data-centre parts from 4NP (Blackwell) to N3P (Rubin) | Observed | Nvidia roadmap | Holds | n/a |
| SA-dwarkesh-2 | Nvidia holds 70%+ of N3 in 2027 | Forecast | SA-silicon-2 | Holds | n/a |
| SA-silicon-6 | Smartphone units fall by low double digits | Forecast | IDC −16.7%, Gartner −8.4%, Omdia −7% | Holds | n/a |
| SA-dwarkesh-3 | Smartphones fall to 500–600 million this year from 1.4 billion | Forecast | Contradicted by IDC, Gartner, Omdia | Doesn't hold | n/a |
Electricity (Malaysia)#
Sources: Gemini result F (GEM-F) and its re-verification (GEM-FR), checked against a published TNB bill and our own June–August 2026 bills. The canonical model is in the site-power doc.
| ID | Claim | Kind | Corroborated by | Verdict | Would be proven wrong by |
|---|---|---|---|---|---|
| GEM-F-1 | RP4 domestic rates: energy 27.03 sen (≤1,500 kWh) or 37.03 sen; capacity 4.55; network 12.85; RM10 retail above 600 kWh | Observed | Both Gemini reports; our bill formula reproduces real bills | Holds | A TNB bill that doesn't fit the formula |
| GEM-F-2 | Above 1,500 kWh the 37.03 sen energy rate applies to every unit (a cliff) | Observed | Reproduces a published bill (1,226 → RM632.72; 1,576 → RM987.55) within RM2–3; the excess-only reading is RM160 out | Holds | Our first bill above 1,500 kWh not fitting |
| GEM-FR-4 | TNB's schedule wording "For all kWh", attributed to the myTNB "understand your bill" page | Observed | That page, fetched directly, carries no tariff table | Doesn't hold (as a citation; the rule itself holds, see GEM-F-2) | A primary document showing the wording |
| GEM-F-3 | 8% sales tax is pro-rated on the share above 600 kWh; the 1.6% levy is on energy + capacity + network above 300 kWh | Observed | Our itemised August 2026 bill, which also adds two refinements: the retail charge is taxed in full, and both the tax and the levy are computed net of the rebate | Holds (refined) | n/a |
| GEM-FR-2 | The efficiency rebate is summed band by band (RM128.50 at 1,000 kWh) | Observed | Contradicted by our itemised August bill: −RM0.025/kWh on all 816 kWh | Doesn't hold. One band rate applies to all units; that reproduces Jun–Aug 2026, with August exact to the sen | n/a |
| GEM-F-4 | Commercial Tariff B becomes cheaper than domestic above about 1,900 kWh | Forecast (arithmetic) | Its own formulas put the crossover at 1,501 kWh (GEM-FR-1 concedes this) | Doesn't hold | n/a |
| GEM-F-5 | Domestic Time-of-Use saves about 1.6 sen/kWh on a flat 24/7 load; the cliff still applies under ToU | Observed | Nothing independent | Pending | TNB's ToU schedule |
| GEM-F-6 | Air-conditioning adds about 0.31 kW per kW of IT heat (COP 3.2) | Forecast | Physics (1 ÷ 3.2) | Holds as a planning figure | Room-level metering |
| GEM-F-7 | At our tier Malaysia is ≈15.6 US¢/kWh: below the US residential average (18.3¢), Europe, Singapore and Japan; about level with Texas and Virginia households; above ASEAN peers and US small-commercial in cheap states | Observed | US residential average also reported by EIA-based trackers | Pending (reported comparison figures) | National statistics that differ |
| GEM-F-8 | RP5 (2028–2030) raises the base 6–10%; our rate follows a base path to about RM0.69 by 2029 (range RM0.56–0.79) | Forecast | Nothing independent | Pending | RP5 consultation papers (expected 2027) |
| GEM-F-9 | Rooftop PV costs RM3–4k/kWp. Panels alone: about RM0.20/kWh levelised, covering about 25% of a 24/7 load, 4.4–4.8-year payback. With batteries: about RM0.40/kWh and about 49% coverage. Solar ATAP caps domestic single-phase at 5 kWac | Forecast | Nothing independent | Pending; no immediate plan (founder, 2026-09-27) | Installer quotes, SEDA rules |
| GEM-FR-6 | TNB allows one account and one meter per residential title; no separate commercial meter without rezoning | Observed | Nothing independent | Pending; moot while the tariff stays domestic | TNB confirmation |
| GEM-FR-7 | Bank Negara's reference rate is about MYR 4.07 per US dollar (late Sep 2026) | Observed | Earlier reports said 4.15–4.25 | Pending | Bank Negara's published rate |
Rental market and demand#
| ID | Claim | Kind | Corroborated by | Verdict | Would be proven wrong by |
|---|---|---|---|---|---|
| SA-rental-1 | H100 1-year rental $1.70 → $2.35 (Oct 2025 → Mar 2026) | Observed | SemiAnalysis's own index (gpu-index-h100); reported by others | Holds (single-source data) | n/a |
| SA-rental-2 | Capacity through Aug–Sep 2026 already booked | Observed | Surveyed providers | Pending | n/a |
| SA-rental-5 | Rental prices more likely to rise than fall | Forecast | SA-dwarkesh-4 | Pending; the index has shown no public update since April | H100 contract index falling below $1.70 while capex guidance still rises |
| SA-dwarkesh-4 | An H100 is worth more today than three years ago | Opinion | SA-rental-1 | Holds for now | As SA-rental-5 |
| SA-rental-4 | Anthropic run-rate $9B → $30B+ in a quarter | Observed | A separate report of $65B run-rate in July | Pending (unaudited) | Lab revenue growth stalling |
| SA-silicon-7 | Google's 2026 capex expectation roughly doubled | Observed | Alphabet guidance of $195–205B | Holds | n/a |
| SA-dwarkesh-5 | Power won't constrain US scaling | Opinion | Contested (a 49 GW shortfall is cited elsewhere) | Pending; doesn't affect our node | n/a |
| SA-dwarkesh-6 | The downside is middling returns on infrastructure | Opinion | Alphabet negative free cash flow; Meta about zero (July earnings) | Holds as the key risk | n/a |