Forward-looking commentary for a general audience, not personalised advice. Prices as-of US close 11 Sep 2026 (StockAnalysis / S&P Global). Research as-of 14 Sep 2026. Horizon: 2–5 years. Voice refresh 17 Sep 2026 (view unchanged).
The market treated “AI infra” as one trade and then sold it as one trade. GPUs get the headlines. Interconnect chips and energised campuses are different businesses with different balance sheets. A 25–43% drawdown from the highs does not automatically mean a broken franchise — or a durable one.
CapEx is huge. The bottlenecks are not the same.
Goldman Sachs Global Institute (1 May 2026) puts a scenario baseline around ~$7.6T of cumulative AI CapEx from 2026–2031, with ~$765B in 2026 rising toward ~$1.6T in 2031 — accelerators, cabling, cooling, and power included. That is a scenario, not consensus.
JLL’s 2026 data-center outlook (5 Jan) sees ~100 GW of new capacity through 2030, up to ~$3T of real estate plus IT fit-out, grid waits over four years in primary markets, and construction costs climbing toward ~$11.3M/MW shell/core — with AI fit-out cited as high as ~$25M/MW.
Keep the constraints separate. GPUs and HBM are the compute unit and a large share of spend. Power and interconnection are often what actually binds deployment. Networking, optics, and ASICs are Astera’s lane — Broadcom’s OFC 2026 stack (Tomahawk 6, CPO, PCIe Gen6) and TrendForce’s July warning that CPO is still yield- and packaging-constrained. Neoclouds and power campuses — Nebius and IREN — consume GPUs; their economics are megawatts, time-to-power, prepay, and financing.
Three claims under one label
Astera Labs (ALAB) sells connectivity semiconductors and COSMOS software for rack-scale AI — PCIe, CXL, Ethernet/UALink, Scorpio fabric, Leo memory controllers, with optical expansion. It is non-GPU: it rides GPU shipments and gets hurt if CPO or in-house serdes shrink the retimer TAM. It is also the only GAAP-profitable name in this trio.
Nebius (NBIS) is a full-stack AI cloud (formerly Yandex N.V.; Nasdaq resumed Oct 2024 as NBIS). Hypergrowth, adj. EBITDA positive, GAAP operating losses, and enormous capex.
IREN (IREN) owns land, grid power, data centers, and GPUs. Contracted AI ARR is scaling fast, but FY26 GAAP revenue was still mostly bitcoin mining.
Reading the “dips”
ALAB is a high-gross-margin chip franchise. NBIS and IREN are capital-intensive capacity stories — trailing sales multiples are noisy; contracted ARR and energised megawatts matter more. Prices below are 11 Sep close.
Astera at $291 (~$50.5B mkt cap) is ~42% below its 30 Jun high of $499 — while Q2 revenue still grew 104% YoY to $392M, GAAP gross margin held 73%, and net income hit $153M. Q3 guide is $540–560M, with Scorpio expected to become the largest product family a quarter early. Liquidity ~$1.25B, no material draw debt, H1 FCF roughly $134M. Concentration is the real scar: four customers at 29% / 25% / 15% / 13% of revenue (some manufacturing partners), with China/Singapore/Taiwan dominating billings. Dilution is modest; valuation is not — trailing P/E ~143x, forward ~52x. This looks like multiple compression after a melt-up, not a disclosed demand break, unless the next prints show customer or Scorpio share loss.
Nebius at $225 (~$62B mkt cap) is ~25% off its high with the guide still intact. Q2 group revenue $582M (+454% YoY), AI cloud ARR $3.0B in June, adj. EBITDA $236M (~41% of sales). FY26 guide: revenue $3.0–3.4B, ~40% adj. EBITDA margin, capex $20–25B, ARR $7–9B, contracted power raised to 5 GW. The balance sheet is a barbell — ~$9.1B cash and restricted against ~$8.5B debt and ~$6B of deferred revenue (prepayments). Named offtake includes Microsoft capacity and a Meta structure around $27B over five years; customer percent of revenue was not disclosed. Dilution is material (ATM ~$2.8B, converts, warrants; shares +22% YoY; short ~20% of float). The debate is financing and delivery, not a vanished logo.
IREN at $44 (~$17B mkt cap) is ~43% off its high. FY26 revenue $707M was still 82% mining; Q4 was the crossover quarter where AI Cloud briefly topped half of sales. Net loss ~$703M was dominated by mining-hardware impairments as sites convert. Management’s ARR language (27 Aug): $1B operating, $4B contracted for 2026 capacity, Microsoft Horizon 1 accepted mid-August. Named contracts include Microsoft ~$9.7B and NVIDIA’s 5 GW partnership plus a $3.4B cloud deal and a 30M-share option at $70. Unrestricted cash $5.9B, debt ~$7.6B, FY27 CapEx guide $25–30B, FY26 FCF roughly −$2.3B. Shares jumped 258M → 394M in about fourteen months (+53%). Short ~27% of float. Horizon 1 is live; the GAAP mix is still catching up — and dilution is a core underwriting input, not a footnote.
What actually separates them
ALAB sells chips that ride every rack. Nebius sells GPU hours and a platform. IREN sells energised campuses still transitioning the P&L off bitcoin. Same “AI infra” sticker; different underwriting.
For ALAB, −42% from the June high with accelerating revenue is mean-reversion after a parabolic print — until concentration or Scorpio share loss shows up. For NBIS and IREN, the falsifier is failed energisation or serial dilution without billed ARR, not a missing Microsoft logo. Only ALAB currently combines fortress liquidity, GAAP profit, and positive FCF. That does not make it cheap at ~140× trailing earnings. It makes it the only name here you can judge on margins rather than slides.
Base case: CapEx continues, interconnect winners keep design-wins, neoclouds raise capital into messy GAAP, mining-to-AI conversions keep writing down old ASICs. Bull case needs Scorpio to become a durable fabric franchise, CPO to arrive slowly enough that copper/retimers stay relevant, Nebius to hit guide without another equity cycle, and IREN to turn $4B of contracted ARR into billed revenue while dilution decelerates. The steelmanned bear is Broadcom/Marvell taking fabric share, CPO collapsing retimer TAM, China/Taiwan CM concentration biting ALAB, power energisation misses, GPU generation rolls stranding utilisation, and IREN’s FY27 $25–30B CapEx forcing another dilutive round.
We break the thesis if ALAB loses a ≥10% customer or misses the Scorpio ramp, if NBIS or IREN cancel material contracts, if energised megawatts miss badly, or if another 20–30% share increase arrives without ARR conversion. Multiples rotating between chips and campuses only changes which sleeve — not whether CapEx exists.
By franchise durability inside 2–5 years: ALAB, then NBIS, then IREN. Asymmetric upside if GW delivery works can elevate the campuses — different question.
Watch ALAB’s Q3 versus the $540–560M guide, customer concentration, Nebius energised MW versus the 800 MW–1 GW 2026 target and ATM usage, Meta/Microsoft tranche deliveries, IREN Horizons 2–4 and AI Cloud’s share of GAAP revenue, FY27 funding versus CapEx, and the CPO / UALink / NVLink narrative.
Bottom line
AI CapEx is real. “Infra dip” is not a single trade. ALAB is a high-quality connectivity franchise sold off after a melt-up — still expensive, still concentrated, still the only GAAP-profit story in this sleeve. NBIS and IREN are GW-scale capacity vehicles where the drawdown is mostly a dilution and delivery debate against contracted logos. We would rather underwrite design-wins and cash generation than LOI megawatts — and we would size the campus names as options on energisation that can be diluted away if the grid and the equity market both tighten.
Primary sources
- Astera Labs Q2’26 PR (4 Aug 2026); 10-Q / 8-K Jun–Aug 2026
- Nebius Q2’26 6-K / shareholder letter (12 Aug 2026); Q1 materials; Nasdaq resume 2024
- IREN FY26 PR (27 Aug 2026); 10-K share count (14 Aug 2026); NVIDIA–IREN PR (7 May 2026)
- Goldman Sachs Global Institute (1 May 2026); JLL Data Center Outlook (5 Jan 2026); Broadcom OFC (12 Mar 2026)
- StockAnalysis / S&P Global (11 Sep 2026 close)