BRAD AI MARKET SIGNAL AI · Semiconductors · Infrastructure · Markets
2026-09-18 · Shanghai Time

Memory Accelerates Across the Board; CoreWeave Proves Compute Pricing Power

Covering the US close of Sep 17 through intraday Sep 18 (data as of 11:23 AM ET, market open).

Today in one sentence: Jensen Huang's "chip sales to double next year" lit up the memory complex, and CoreWeave's ~$40M/MW short-term contracts proved compute pricing is still rising — but NAND price hikes are already decelerating, and the cycle is entering the verification-heavy middle innings.

01 — Top Signals

1. Memory: "sales doubling" + price-hike resonance — the supercycle is being priced for its acceleration phase

What happened. On 9/17, Jensen Huang told reporters in Scotland that "NVIDIA will sell twice as many chips next year as this year" (unit count, not revenue — CNBC first, Bloomberg corroborated). The same day, Reuters exclusively reported that SK hynix is in talks with Intel about producing memory chips in the US for the first time. The two catalysts combined sent memory sharply higher: MU +5.5% to close at $977.5, SNDK +6.2% to close at $1,614 (up another +6.1% intraday on 9/18), and SK hynix closed +5.1% in Korea on 9/18.

Why it matters. The transmission chain is clear: GPU unit sales doubling → synchronized upward revisions to HBM4/DRAM/NAND bit demand → 2026 HBM capacity already sold out (per company filings), 2027 essentially sold out (per supply-chain checks) → price and share become the only remaining variables. The SK hynix–Intel talks are the incremental piece: if consummated, memory makers are beginning to lock in overseas capacity for 2027–28 demand, extending expectations for cycle length.

Investment implications. Near-term pricing is already quite full (MU is only 8 trading days from earnings; SNDK is up roughly 12% over two sessions), and crowded positioning into the 9/30 MU print is the main risk. The medium-term debate is not whether the shortage persists, but the slope of the increases. Falsification conditions: MU guidance on 2027 contract prices softening, or TrendForce's 4Q26 forecast showing a marked convergence in price gains.

HIGH IMPACT Confidence: HIGH Affected: MU · SNDK · SK hynix · Samsung · NVDA

2. CoreWeave: $40M/MW short-term contracts prove pricing power — but a same-day $3B debt raise plus equity offering decided the stock

What happened. CoreWeave disclosed: newly signed 3–6 month short-term contracts in Q3 priced at roughly $40M/MW (annualized); contracted power increased from 3.7GW to about 4.2GW; over $25B in net new customer commitments added early in Q3. The same day, it announced a $3B convertible notes offering (due 2033) and an ATM program of up to 35 million shares (8-K filed). The stock closed down 4.2% on 9/17.

Why it matters. $40M/MW is a scarce front-line price data point: rising short-term contract pricing directly falsifies the "compute deflation" narrative, and Nebius the same day notified customers of higher on-demand GPU rates effective 10/1 — neocloud pricing power is industry-wide. But the market voted with a decline: the faster the expansion, the deeper the dependence on external capital. An FT long-form piece the same day called out the OpenAI–MSFT–ORCL–NVDA circular deal structure, and a backlog-quality discount is beginning to be discussed.

Investment implications. Both bulls and bears have hard evidence; the net call hinges on financing costs. The convertible pricing (coupon and conversion premium, landing in the coming days) is a direct read on neocloud cost of capital. If the coupon comes in meaningfully above expectations, the "GPU leasing is a great business" ROIC narrative gets re-rated. Falsification condition: pricing power fails to convert into positive free cash flow.

HIGH IMPACT Confidence: HIGH Affected: CRWV · NBIS · IREN · NVDA

3. Google seeks a CoWoS backup for its 2027 TPU: Intel EMIB-T yields lag, TSMC packaging gets tighter

What happened. DigiTimes reported on 9/17: Google's next-generation TPU (ASIC partner: MediaTek), slated for volume production in late 2027, was originally set to use Intel's EMIB-T packaging, but with substrate yield ramp lagging (validation yield ~90% versus a ~98% production threshold), Google is evaluating a backup shift to CoWoS. The same day, TrendForce published: CoWoS-L, on maturity and yield, will remain the mainstream advanced packaging for AI chips through 2028.

Why it matters. A two-way signal. For TSMC: if the Google backup materializes, CoWoS demand tightens further; TSMC's August revenue already set an all-time monthly record (NT$514.8B, +53.3% YoY, still accelerating sequentially), extending the packaging-capacity premium cycle. For Intel: EMIB-T is the core chip in its foundry-rebound narrative, and hesitation from a marquee customer hurts the valuation logic more than a lost order.

Investment implications. Evidence grade is supply-chain media (DigiTimes), unconfirmed officially — treat position sizing as an "option," not a "fact." Falsification conditions: Google/Intel confirm EMIB-T is on schedule, or MediaTek issues a clarification.

MEDIUM IMPACT Confidence: MEDIUM Affected: TSM · INTC · GOOGL · MediaTek

02 — Market Moves That Matter

Context: On 9/16 the Fed raised rates 25bp to 3.75–4.00% (the first hike under Chair Warsh; the dot plot hints at possibly one more this year). On 9/17, fading uncertainty drove a rebound: Nasdaq +1.69%, SOXX +3.39%. 9/18 is triple witching — late-session volatility will be amplified by options structure.

SNDK+6.2% (9/17) · +6.1% (9/18 intraday)
Primarily fundamental (NAND tightness + JPM price target of $2,250); but triple witching plus large $1,600 call positioning (~$41M notional) means part of the move is gamma amplification — poor risk/reward for chasing.
MU+5.5% (9/17) · +1.3% (9/18 intraday)
Fundamental: the Sanand, India assembly/test plant came online + Huang's sales-doubling comment + building anticipation for the 9/30 print. Heavy call positioning (~$44M notional) sits ahead of the $1,000 strike; triple witching could affect the short-term path.
AMD+6.4% (9/17) · -0.5% (9/18 intraday)
No company-level news of sufficient magnitude to independently explain the move (the SiFive ROCm collaboration is too small). More likely sector beta + catch-up + momentum flows after a three-day run. We don't invent stories.
INTC+7.7% (9/17)
SK hynix talks rumor + analyst target hikes, but call volume/OI hit 362x — positioning weight is very high. The stock is up roughly 2x YTD; the gap between valuation and foundry execution progress is widening.
CRWV-4.2% (9/17)
Fundamental/Financing: dilution concerns from the $3B convertible plus 35-million-share ATM offering outweighed the $40M/MW pricing positive. The market is re-pricing neocloud capital intensity.
ORCL-3.4% (9/18 intraday, after +5.2% the prior day)
Recurring media reports on Ellison's share-sale plan stirred discussion; give-back in nature. No evidence of fundamental deterioration.

Others: AVGO +2.9% (9/18 intraday; Rosenblatt raised its target $500→$600, the Street high); AMAT +3.8% / LRCX +4.6% (9/18 intraday; equipment catch-up, tracking expectations for a TSMC CapEx raise); COHR +4.8% (1.6T optical-module narrative continues; no company-level news).

03 — Semiconductor Signals

  • The slope of HBM price increases is converging (FACT). TrendForce's current 3Q26 read: Conventional/Server DRAM contract prices +13–18% QoQ (versus +58–63% in 2Q26); NAND +10–15% (versus +70–75% in 2Q26). Direction is still up, but acceleration has clearly slowed. The 4Q26 forecast is expected around 9/24–25 — the next hard data point.
  • Kioxia voluntarily paused NAND price hikes (FACT). The CEO told Bloomberg that "raising prices too far would weaken data-center customers' ability to invest." This is supply-side speed control, not a demand collapse — but it means the steepest leg of NAND spot pricing is likely behind us.
  • CXMT enters NAND (Reuters exclusive). A new Beijing fab will build NAND R&D and production lines, competing head-on with YMTC; mass-production timing undetermined. No near-term impact, but the China variable on the 2027–28 supply side is starting to accumulate. A second Reuters exclusive the same day: Solidigm is evaluating a NAND fab in New York State (no contract signed).
  • The Broadcom $230B anchor holds. Hock Tan reiterated on CNBC on 9/14 the FY2028 AI revenue guidance of $230B (FY2026 expected ~$58B, +186%), saying supply is locked. Note: AVGO fell 4.8% on the day of the reiteration — the market's pricing of the ASIC narrative is already very full, and Rosenblatt's $600 target (9/17) is sell-side chasing, not new information.
  • The packaging mainline confirmed. TrendForce 9/18: CoWoS-L remains mainstream through 2028; 2026 high-end GPU shipments estimated +30% YoY. Combined with TSMC's record August revenue, the "AI CapEx has peaked" thesis continues to be falsified by the data.

04 — AI Infrastructure

  • The bottleneck continues migrating toward power (FACT). Amazon signed a long-term backup-generator agreement with Generac worth up to $8B (first phase $2.4B, deliveries 2027–28, including warrants); Generac rose as much as ~40% in a single session. The backup-power supply chain is the next trade in the power bottleneck. The same day, the US House passed the Ratepayer Protection Act 417–3 (data centers ≥150MW must self-fund grid upgrades); it has stalled in the Senate for now — policy risk is real but not yet binding.
  • Rubin ramp cadence gets empirical confirmation. CoreWeave brought a multi-rack Vera Rubin NVL72 cluster online on 9/16 (hundreds of Rubin GPUs in a single scale-out cluster, per the official press release); SemiAnalysis estimates Vera Rubin's profit per GW at 2x Blackwell's. The unit-economics improvement of the Rubin generation is an important pillar supporting 2027 GPU demand expectations.
  • OpenAI in talks for a round at a $1.2T valuation (FT first, Reuters/WSJ followed). This is not just a valuation story: OpenAI's solvency directly determines the credibility of its hundreds of billions in compute-purchase commitments to Oracle/Microsoft/CoreWeave. The FT long-form the same day mapped the circular deal structure; the market is beginning to discount backlog quality.

05 — What the Market May Be Missing

1. The market is trading "memory shortage persists" — and may not be fully pricing "slope convergence." DRAM contract-price QoQ gains have converged from +58–63% to +13–18%, NAND from +70–75% to +10–15%, and Kioxia voluntarily paused hikes — classic mid-cycle characteristics. Yet the stocks are accelerating (SNDK +12% in two days). If TrendForce's 4Q26 forecast (~9/24–25) shows further convergence, or MU's 2027 guidance on 9/30 is vague, the "supercycle" trade will see violent disagreement. This is not a bearish call — it means current prices depend more on "acceleration continuing" than on "the shortage existing."

2. A possible Rubin Ultra spec cut: the demand-side crack in the HBM bull narrative that almost nobody is pricing. SemiAnalysis and The Information previously reported that NVIDIA, constrained by HBM supply, is testing a de-spec'd Rubin Ultra (the mainline SKU may settle at 192GB HBM4, a large reduction from the original plan); some analysis suggests 2027 HBM demand could be ~10% lower as a result. NVIDIA denied only a platform delay, not the spec cut. The market trades the HBM shortage as a pure supply story; if the shortage is real enough to force NVIDIA to de-spec, the demand-side numbers themselves need downward revision — a two-way risk that is barely being discussed.

3. Neocloud pricing power and cost of capital are two ends of the same chain. The $40M/MW short-term contracts and NBIS rate hikes prove the revenue side, but CRWV's same-day debt plus equity raise — met with a -4% response — shows the pricing function has switched from a "demand story" to a "capital-structure story." The coupon on the upcoming convertible will do more to determine the valuation center of CRWV/NBIS than any bullish research note.

06 — Earnings / Estimate Watch

  • MU (reports 9/30): Street expects FY26 revenue of roughly $41.5B. Three must-watch items: HBM4 shipments and share (relative position vs. SK hynix/Samsung); 2027 contract-price guidance (determines whether slope convergence continues); CapEx and buyback scale (the capital-discipline signal at a cycle top). Options are crowded around the $1,000 strike into the print, and implied vol is priced rich.
  • AVGO: LSEG FY2028 EPS consensus is $25.86, while the company's $230B guidance implies >$30 — if the supply chain keeps validating (Anthropic's 5GW of TPUs in 2027, OpenAI's Jalapeño in production by year-end), consensus has structural upside. This is one of the clearest expectations gaps among large-cap AI names today.

07 — Catalyst Calendar (next 7–14 days)

9/18Q3 triple witching — late-session volatility amplified; watch quarter-end rebalancing in the big winners (MU +256% YTD, INTC +225%)
~9/24–25TrendForce 4Q26 DRAM/NAND contract-price forecast (inferred from the historical cadence) — the next hard data point for memory
9/25US August core PCE — the first inflation read after the hike
9/30Micron FY26Q4 earnings (after the US close) — memory's most important September catalyst
10/1Nebius's new GPU rental rates take effect (pricing-power validation point); CoreWeave convertible pricing window

08 — Radar

  • SK hynix × Intel US memory production: after the Reuters exclusive, the company responded "exploratory, not finalized"; the Korean government may oppose HBM production in the US. Follow up next week.
  • Apple enterprise AI servers: The Information reports Apple is developing an M8 Ultra inference server and discussing adoption of NVLink Fusion; if real, the NVLink ecosystem gains a top-tier customer (positive for NVDA/AVGO). Still only at the "discussion" stage.
  • Anthropic IPO: the underwriting syndicate is targeting October, potentially beating OpenAI to market; watch for the compute-cost structure in any S-1 leak.
  • Crusoe's $3.9B raise: OpenAI's largest data-center builder is pivoting to factory-prefabricated small data centers — a marginal change in construction model worth tracking.
  • NVIDIA NVHBM custom HBM: Samsung is already a core partner (8-layer HBM4E @17–18Gbps); SK hynix's reliance on external foundries for base die is a relative disadvantage and a medium-term share variable. Not a new event — ongoing tracking.

09 — Bottom Line: 3 Things to Watch

  1. The 9/30 MU print — HBM4 share and 2027 contract-price guidance. This is the watershed between "accelerating" and "converging at a high level" for the memory cycle, and a stress test for today's crowded positioning.
  2. TrendForce's 4Q26 contract-price forecast (~9/24–25) — if DRAM gains keep converging while NAND flattens, the acceleration premise of the "supercycle" trade needs re-rating.
  3. CoreWeave's convertible pricing + NBIS's 10/1 rate hike taking effect — the net effect of neocloud cost of capital and pricing power will determine how fast the CRWV/NBIS valuation center switches from a demand story to a capital-structure story.

Sources & Further Reading

  • CNBC — Jensen Huang: Nvidia to sell twice as many chips next year
  • CoreWeave IR — New compute contracts at ~$40M/MW
  • SEC 8-K — CoreWeave $3.0B convertible notes + ATM program
  • TrendForce — CoWoS-L to remain mainstream advanced packaging through 2028
  • DigiTimes — Google weighs CoWoS backup for 2027 TPU as EMIB-T yields lag
  • Reuters — SK hynix's Solidigm weighs US NAND plant; SK hynix–Intel US memory talks; CXMT enters NAND
  • Reuters/Yahoo — Nebius hikes AI cloud GPU rental rates from Oct 1
  • MarketWatch — Generac–Amazon up to $8B generator deal
  • Financial Times — Joining the dots between big AI
  • WSJ — OpenAI weighs pre-IPO round at >$1.2T valuation
  • TSMC — August 2026 revenue NT$514.8B, +53.3% YoY, record high
  • Federal Reserve — FOMC raises 25bp to 3.75–4.00% (9/16)
  • CoreWeave — Multi-rack Vera Rubin NVL72 cluster online
  • The Information — Nvidia weighs less memory for Rubin Ultra