BRAD AI MARKET SIGNAL AI · Semiconductors · Infrastructure · Markets
2026-09-19 · Shanghai Time

Neocloud Capital Day: Nscale Sprints Toward a $35B IPO; CoreWeave Locks In 2.875%

Covering the full US session of Friday Sep 18 (quad witching close) through Saturday morning Asia/Shanghai.

Today in one sentence: Nscale filed for a NYSE IPO targeting up to $35B with a $44.6B Anthropic contract in hand, and CoreWeave priced an upsized $3.7B convertible at just 2.875% — the neocloud capital race is accelerating; memory stocks melted up into quad-witching, with MU closing above $1,000 for the first time.

01 — Top Signals

1. Nscale files for a NYSE IPO: the prospectus is the industry's first public physical exam

What happened. On 9/18, Nvidia-backed British AI cloud/data center company Nscale publicly filed for a NYSE listing (SiliconANGLE relaying the prospectus; Reuters/FT/WSJ/Bloomberg all confirmed): it aims to raise up to ~$3B (per Bloomberg) at a targeted valuation reportedly as high as ~$35B — a ~2.4x re-rating in six months from its $14.6B March private mark. Key prospectus numbers: H1 2026 revenue of $140.6M (+1,250% YoY); net loss of $1.02B (nearly 3x YoY); last month's $44.6B infrastructure contract with Anthropic (~460MW at the Monarch campus in West Virginia, microgrid-powered); Figure AI committed to renting up to 100,000 GPUs (Nscale taking a stake in Figure); plus a multibillion GPU hosting agreement with Microsoft. It has 25,000 GPUs deployed and 461,000 "active and contracted," claims a "clear expansion path to 8GW+," recently issued $3.1B of convertible notes (at least $1B taken by Nvidia), and has raised $4.6B+ of debt YTD.

Why it matters. Three readings: ① Price calibration — $44.6B/460MW ≈ ~$97M of full-cycle contract value per MW, which cross-checks against CRWV's $40M/MW annualized short-term contracts: long-term lockups are cheap, spot/short-term is expensive, and the spread itself prices scarcity. ② Capital structure — Nvidia graduates from supplier to creditor (≥$1B convert), one more rung in the circular-financing stack. ③ Valuation yardstick — if the $14.6B→$35B re-rating lands in public markets, CRWV/NBIS get their first same-generation comp; if it doesn't, it's a stress test of the sector's liquidity.

Investment implications. For CRWV/NBIS: more valuation anchors are good, but AI-infra paper supply is arriving in the same week — whether the marginal buyer is thick enough is the core question. The prospectus's 18x gap between "25k deployed vs 461k contracted" is the industry's capacity leverage in miniature: fulfillment needs CapEx, CapEx needs financing — the chain is intact but every link is getting more expensive. Falsification: IPO priced well below the $35B target or postponed, or GPU depreciation-cycle assumptions publicly challenged by sell-side.

HIGH IMPACT Confidence: HIGH (multi-source) Affected: CRWV · NBIS · NVDA · TSM · Anthropic (private)

2. UPDATE — CoreWeave's convert prices at 2.875%, upsized to $3.7B: the cost-of-capital question gets its answer

What happened. (UPDATE — the increment on yesterday's story) On 9/18 CoreWeave priced its convertible notes: upsized from $3.0B to $3.7B, plus a $500M option for initial purchasers (up to $4.2B); coupon 2.875% (due April 1, 2033, settling 9/22); initial conversion price $97.85, a 22.5% premium to the 9/17 close of $79.88; $498.8M spent on capped calls with a $199.70 cap (150% premium); net proceeds ~$3.64B (terms per the pricing release; structure per the SEC 8-K exhibit). The stock closed +1.85% at $81.36 on 9/18.

Why it matters. Yesterday's open question — "the coupon is the direct read on neocloud cost of capital" — now has its answer: 2.875% is far below CoreWeave's existing straight debt at 8.5–9.75% and close to its older converts issued in 2024–25 at 1.75%. In other words, credit markets still fund CoreWeave on "high-vol growth" terms, not "distressed credit" terms — demand was strong enough to upsize by $700M. $40M/MW of pricing power × 2.875% funding cost: that spread is currently the hardest data point supporting the GPU-leasing ROIC narrative.

Investment implications. Financing-cost fears are temporarily resolved, but the total quantity of dilution hasn't changed: up to $4.2B of converts plus a 35M-share ATM still overhang the stock — only the price and pace changed. Watch: the 9/22 settlement and whether the $500M option is exercised; whether the stock can hold near the $97.85 conversion price — the $199.70 capped-call cap implies structural hedging flows in the $97.85–$199.70 band. Falsification: a deep break below the conversion price, or materially wider spreads on the next raise.

HIGH IMPACT UPDATE Confidence: HIGH (company release) Affected: CRWV · NBIS · IREN · NVDA

02 — Market Moves That Matter

Context: quad-witching close. S&P 500 finished at 7,650.50 (+0.17%), Nasdaq +0.5%, SOXX +2.69% — semis stood alone on options expiration. Rotation out of mega-cap platforms into hardware: META -2.43%, MSFT -0.80%, ORCL -1.98%. Week in review: after the 9/16 Fed hike (to 3.75–4.00%), two straight rebound sessions.

SNDK+10.99% to $1,791.82 (+17.9% over two days)
No single company-level news. Fundamental (memory repricing) + Positioning (Seeking Alpha: recovery after two down weeks), amplified by witching-day gamma.
MU+3.92% to $1,015.80 — first close above $1,000
Fundamental: pre-earnings (9/30 after close) positioning; Q4 guide $49–51B revenue / $30–32 EPS. Options imply an ±11% earnings-day move — the market trades the "beat magnitude," not the shortage itself.
AMAT/LRCX+6.51% / +6.98% (ASML +3.08%, KLAC +4.74%)
No standalone company news — continuation of TSMC CapEx-raise expectations. Equipment is the most sensitive to "2027 CapEx peaks vs extends"; the market is clearly betting on the latter.
COHR/CRDO/LITE+7.22% / +4.54% / +4.17%
1.6T optics momentum continues; no fresh company-level catalyst. Sector beta + expiration flows; no story-telling.
GEN-3.97% (give-back after the Amazon-deal surge)
Classic "deal announced → profit-taking" path; separately, reports say Bloom Energy is joining the S&P 500 yet still fell 5.4% Friday.

Others: NVDA +1.34% ($222.27), AMD +2.70%, AVGO +2.97%, INTC -0.18%, TSM +1.02%, MRVL +1.45%, CRWV +1.85%, NBIS +2.55%, IREN +7.36%, AMZN +1.00%, GOOGL +0.64%, ANET -0.07%.

03 — Semiconductor Signals

  • AMD posts first official EPYC "Venice" benchmarks, aimed at Nvidia Vera (FACT, vendor claims). On 9/18 AMD published the first official benchmarks for EPYC "Venice" (Zen 6, 256-core): claiming system-level performance more than 2x Nvidia's 88-core Arm Vera, with the 96-core SKU +20% per-core (Tom's Hardware). Background: Vera is expected to generate ~$20B revenue this fiscal year (sell-side estimates). Vendor-run benchmarks deserve a haircut, but the signal is clear: the fight over CPU share of the AI-server BOM has begun.
  • SK hynix × Intel (UPDATE, UNCONFIRMED). Follow-up reporting points to an Intel Ohio site as the potential landing spot; talks remain exploratory, and the Korean government's stance is the swing variable.
  • Memory prices: next hard data point is TrendForce's 4Q26 forecast (~9/24–25). With stocks priced for slope rather than level, the marginal change in that forecast matters more than the absolute print.

04 — AI Infrastructure

  • The bottleneck migrates from "signing mega-deals" to "grid-connection speed" (CNBC exclusive). CNBC reported Anthropic and OpenAI are hunting for 20–30MW-class smaller data center deals in the UK and Nordics to get power and compute faster. Read: power queues and construction timelines at mega-campuses are now the binding constraint; labs are paying a premium for time — the same phenomenon as CRWV's $40M/MW short contracts and Nebius's 10/1 rate hike. What's scarce is no longer land and blueprints; it's "racks that can be energized now." Bullish for neoclouds with live capacity and for existing powered DC pricing.
  • Policy risk extends from federal to state (FACT). Virginia Governor Spanberger on 9/18 signed a data-center "accountability" executive order and created an AI task force (WSLS), stacking on the House's Ratepayer Protection Act — regulatory tightening in the world's largest data-center state is a slow-burn risk for developers with heavy VA exposure.
  • Nscale prospectus CapEx detail. $1.85B for a 200MW campus in Ward County, Texas (some racks with Nvidia Rubin GPUs and Vera CPUs), plus $1.2B to upgrade a North Carolina site — note that Rubin systems are now showing up in non-hyperscaler order books and financing use-of-proceeds; Rubin demand "breadth" is being validated.

05 — What the Market May Be Missing

1. Neocloud paper supply is arriving in a single week — nobody has tested the marginal buyer's depth.

CRWV up to $4.2B (convert + ATM) + Nscale's ~$3B IPO + Nscale's $4.6B+ YTD debt and $3.1B convert — the same pool of AI-infra risk appetite is being asked to absorb multiple tranches at once. Every deal so far has priced (CRWV even upsized), but "price hasn't broken" ≠ "quantity isn't a problem." Nscale's IPO pricing is the first public stress test of this demand.

2. The MU print trades the "beat magnitude," not the shortage itself.

The Q4 guide ($49–51B revenue, $30–32 EPS — roughly 10x a year ago) has been public for months; options imply ±11%. A beat is nearly zero surprise; the real variables are the 2027 contract-price language and the HBM4 share narrative. Meanwhile SNDK's +11% was a squeeze-style repair after two down weeks — positioning was amplified by witching, and volatility cuts both ways.

3. "Time-to-power" is becoming an underpriced variable.

Labs pivoting to 20–30MW sites says power-queue time is now the hard constraint on compute supply. The essence of $40M/MW short contracts is a queue-jumping fee — it prices "now," not GPUs. Watch: the share of early-energization clauses/penalties in new neocloud contracts, and when the valuation gap between already-powered existing DCs and new campuses starts to widen.

06 — Earnings / Estimate Watch

  • MU (9/30 after close, FY26Q4): guide of $49–51B revenue (~+345% YoY) and non-GAAP EPS $30–32 (company guidance via a Motley Fool preview). Three must-watch items: HBM4 shipments/share, 2027 contract-price language, CapEx/buyback discipline. With the 10Y near 5%, high-beta cyclicals are more sensitive to Fed rhetoric than usual.
  • AVGO (continuing watch): the gap between FY2028 consensus EPS ($25.86) and what the $230B AI-revenue guide implies (>$30) still hasn't closed — needs earnings-season confirmation.

07 — Catalyst Calendar (next 7–14 days)

9/22 (Mon)CRWV convert settlement + decision on the $500M option; watch Nscale roadshow progress (reportedly as early as September)
~9/24–25TrendForce 4Q26 DRAM/NAND contract-price forecast — next hard data point on the memory slope
9/25US August core PCE — first inflation print after the hike
9/30Micron FY26Q4 earnings (after close) — options imply ±11%
10/1Nebius new GPU rental rates take effect (pricing-power check)

08 — Radar

  • NVDA anchoring Anthropic's IPO (mind the dates): reports of Nvidia taking up to $10B at a ~$2T target valuation cluster around 9/11–17 (Yahoo summary), recirculated 9/19; IPO reportedly targeted for October.
  • Trump–Xi White House state dinner (9/19): Nvidia and OpenAI CEOs among attendees — a live two-way risk event for AI-diffusion and export-control policy.
  • Anthropic revenue run-rate rumor: some social accounts claim annualized revenue above $100B — evidence grade too weak (Tier 5); awaiting confirmation.
  • SK hynix × Intel Ohio (UNCONFIRMED): exploratory talks; the Korean government's stance is the biggest variable.

09 — Bottom Line

Brad's 3 Things to Watch
  1. Nscale IPO pricing vs the CRWV/NBIS comp — the first public stress test of marginal AI-infra demand; a discount or delay would be a sector-level signal.
  2. MU on 9/30 (±11% implied) — the 2027 contract-price language and HBM4 share narrative matter far more than the size of the Q4 beat.
  3. TrendForce 4Q26 forecast (~9/24–25) — DRAM hikes converging or extending decides the direction valve for crowded memory positioning.

Sources & Further Reading