Good morning. CXMT IS THE TAPE. Memory orders stretch through 2027, PC vendors racing to lock supply, and this IPO is pricing in full-on mania. Nomura slaps a RMB 7.76T market cap target on an $85B IPO — 13x from the offering. No one cares about mean-reversion arguments. They're just wrong.
Futures mildly lower. S&P off 0.3%, Nasdaq -0.5%. Mag7 spenders getting repriced (-5.8% last week) while semis held -1.2%. The ROIC debate is pulling the tape, not fundamentals. AH prints are quiet — no major earnings bombs. The CXMT listing is the real event.
Asia: China opens, CXMT begins trading in 3 hours. This will be the craziest IPO I've ever seen. Oil stays elevated on Iran — JPMorgan flags $114/bbl Brent on extended disruption. That's a risk bid, not a growth bid.
Three themes frame the day:
1. Memory is the structural AI bottleneck, full stop. The Micron/Meta whitepaper quantified it: 38x slowdown when DRAM spills to SSD. CXMT's IPO validates the thesis at extreme multiples. Bear cases are shallow — mean reversion doesn't apply when AI goodput depends on HBM capacity. MU, SKHY, CXMT own this.
2. Power is more supply-constrained than chips. GEV's gas turbine backlog hits 116 GW with delivery into 2031. This is the binding node. Less exposed to credit cycles or model debates. ADI's 73% gross margins reflect design-in moats — but China analog competition is the surveillance risk for new wins.
3. Vertical integration is accelerating across the stack. INTC/SK Hynix Ohio JV confirmed. AMD adopts CPO for MI500 in 2027 (likely Ayar Labs). Optical interconnect inflection validates networking layer. Foundry partnerships are real — not just slideware.
We'll hit up CXMT and GEV first, then get to ADI and the memory complex.
This week's Advancing AI event was a forward-operating-base showcase, not a Q2 catalyst. The headline numbers are staggering — AMD raised its 2030 TAM to ~$2T, with data center accelerators alone hitting ~$1.4T (45%+ CAGR). They disclosed 14 GIGAWATTS OF COMMITTED COMPUTE, which Stifel pegs at ~$245B of cumulative revenue at ~$17.5B/GW. That's a 3-4x multiple on current annual revenue just from committed orders. The narrative is clear: the order book is no longer aspirational, it's contractual.
AMD is shifting to an annual rack-scale cadence (Helios/MI455X in prod now, 500/600 in '27/'28). They announced Helios is in full production — first in-house rack system. That's a big deal for capturing hyperscale wallet share. The EPYC Venice (Gen 6) is also in full production, supporting the server CPU TAM raise to ~$220B.
The partnership story is the real muscle: Anthropic deal (potentially $17B), Cerebras for disaggregated inference, Kia robotics platform, ROCm.ai. They're building a full-stack narrative to compete on open standards vs. NVDA's moat.
Stifel stick with Buy/$635 and called the >$20 adjusted EPS target a "conservative floor":
"The 14 GW committed compute figure represents approximately $245 billion of cumulative revenue based on our framework of roughly $17.5 billion per gigawatt. We view the company's longer-term adjusted earnings per share target exceeding $20 as a conservative floor."
Bull case: Revenue growth is real — 43% for FY26. The multiple compresses if they execute. Baird took it to $1,250 (wild) on a 15% share of $147B AI GPU by 2030. Melius to $660 on the Anthropic deal alone. At 178x P/E, the stock needs this growth to arrive.
Bear case: William Blair makes the right point: valuation embeds "increasingly high expectations with little room for error." At $856B market cap and $7.50 EPS for FY26, you're paying for 2028/2029 earnings today. One execution miss, one hyperscale pushback, one NVDA competitive response that narrows the moat — and the multiple compresses hard.
This is a positioning debate for PMs, not a fundamental one. The revenue trajectory is real. The TAM is enormous. The customer concentration (hyperscalers) and competition (NVDA's software moat) are the real bogeys. At 175x trailing, the stock needs both the revenue ramp and multiple compression to work simultaneously. That's a high-conviction wager. The event strengthened the bull case long-term, but near-term r/r still feels stretched unless you're building a 3-year position.
PEGA is getting obliterated — down 19% in a week, sitting at 52-week lows after a Q2 miss fueled by AI confusion freezing customer decisions. Revenue $420.7M (missed by $7M), EPS $0.35 vs $0.43. Cloud revenue grew 28% y/y to $213.9M but still short of estimates. The bull case? 12% FCF YIELD AND A PEG OF 0.25 — obscenely cheap if the AI paralysis is transitory. DA Davidson cuts PT to $32 from $55 but stays Buy, pointing to FCF upside in FY27. Others are less forgiving: Loop downgrades to Hold ($25), William Blair to Market Perform. Rosenblatt keeps Buy but trims PT to $47. Risk/reward tilts deep value, but the near-term narrative is horrid.
"The company’s strong 12% free cash flow yield and low PEG ratio of 0.25 suggest the stock may be undervalued at current levels." — DA Davidson analyst Lucky Schreiner
Data center optical interconnects are the only story that matters here — and it's a monster. Stifel goes to $120 (was $110), Needham to $100 (was $60), both Buy. The June quarter revenue of $168.8M cleared Stifel's bogey by 2.3%, but the real signal is the September guide: midpoint $215M, up 27% QoQ and a staggering 22.9% above their prior estimate. That's not seasonality — that's a step-function ramp.
The optics narrative gets bigger every print. MXL raised CY26 optical data center revenue guidance to $210-230M from $150-170M. Keystone is already ramping across 800G deployments at multiple major customers in both the U.S. and Asia, and the 1.6T platform is sampling for late CY27 revenue. At $120, Stifel's target implies 47x CY27 P/E — that's aggressive, but the growth rate (55% YoY) and order of magnitude shift in scale (from <$200M to >$800M revenue in two years) make the multiple less scary.
"Keystone is ramping across 800G deployments at multiple major customers in the U.S. and Asia."
The after-hours dip post-earnings was noise — growth sustainability concerns are overblown when you have 800G and 1.6T product cycles back-to-back. This is a three-year compounder in a market that's still early innings. (Price action to watch: stock at $75.97, nearly 50% below the Stifel target — implies the market still discounts the optical ramp. That's the opportunity.)
Verdict: The data center fiber opportunity is the primary leg of the bull case – and UBS's model just validated Dycom's own $20B five-year estimate. This isn't a legacy telco contractor anymore. This is an infrastructure beneficiary of the AI data center buildout. The Q1 earnings beat (EPS $4.42 vs $2.72, revenue $1.965B vs $1.67B) only reinforces the momentum. Stock up 68% over the past year, but the opportunity is still early.
UBS lays out the math: a $19B market opportunity for contractors over 2025-29, with annual spending nearly tripling from ~$2.3B to ~$6B. Their model projects $14B of that falling in 2027-29. Dycom's own $20B estimate had upside caveats – UBS says the model validates that number.
"Data center fiber connections have emerged as one of the key growth pillars for Dycom and other telecom contractors over the next several years."
The analyst crowd is consolidating around this thesis. UBS, KeyBanc, and Cantor Fitzgerald all raised PTs to the $610-654 range following the Q1 blowout. The fundamental driver is fiber-to-the-home strength and margin expansion – but the real second-derivative story is data center fiber. Revenue growth of 30% LTM and 37% expected for FY2027? That's top-tier in the space.
The risk: This is an execution story in a capex-heavy industry. If hyperscaler spend slows or fiber construction timelines slip, the multiple compression can be violent. But the rate of change is still strongly positive, and Dycom keeps delivering. For now, the bull case wins.
Verdict: Range-bound until new leadership materializes. UBS cut PT to $9 (from $10), staying Neutral as the CEO transition injects uncertainty into an already messy pivot toward owning robotaxis. Mizuho follows with $8, same Neutral. Canaccord still Buy at $12 — but that's down from $17, acknowledging competitive pressure from Chinese players and limited VW traction.
The Q2 beat ($0.19 vs $0.06, rev $508M vs $482M) is overshadowed by Shashua stepping down after 12 years. Co-founder stays on board, but the model shift from AI/software to physical operations (robotaxis, humanoids) is a different execution game. The Israeli R&D tax credit helps cash flow near-term — but larger proof points (customer clarity, new CEO vision) are absent.
"The firm will remain on the sidelines, believing the stock could be range bound until larger proof points and new leadership emerge." — UBS
MBLY remains a show-me story. Earning power intact, but narrative needs a new face. PMs should wait for the CEO search to narrow before leaning in.
Verdict: Show-me stock. Benchmark reiterates Buy with $525 PT, but that's barely above the $383 tape — the market is pricing in serious execution risk. The real debate is whether Azure growth and Copilot monetization can justify the massive AI capex. July 28 F4Q26 print is the next catalyst, and everything hinges on the rate of change in Azure demand and free cash flow margins.
Benchmark's strongest narrative point — the shift from frontier AI partner to AI operating system layer — is the long-term bull case. But near-term stock reaction is all about fundamentals: revenue growth, operating margin, FCF. Other shops (Bernstein $646, Truist $575, Oppenheimer $515) all on the same side, but consensus doesn't move the needle until the numbers prove it.
"Microsoft is transitioning from a frontier AI partner into an AI operating system and orchestration layer for enterprise AI adoption." — Yi Fu Lee, Benchmark
That strategic pivot is real, but PMs should watch for any sign of deceleration in Azure or softer Copilot seat uptake. The bull case works if the rate of change stays positive.
Stifel cuts PT to $110 from $120 (Hold), saying the 172% YTD rally has already priced in the early turnaround. The Q2 beat was massive — revenue $16.1B (+25% YoY, $1.8B above guide), EPS $0.42 vs $0.22 consensus, gross margin 41.8% (+280bps). AI-driven businesses now 70% of revenue. But the most important catalyst for the next leg — a signed external foundry customer for 14A — hasn't arrived.
"The 2026 re-rating has already priced a meaningful portion of the early turnaround... the most important 14A catalyst, a signed external foundry customer, has not yet arrived."
Rest of the Street is more mixed. DA Davidson raised to $100 (Neutral), Truist to $108, TD Cowen to $115 (Hold), Baird to $125. Rosenblatt slashed to $80. Consensus seems to be "great quarter, but how much further?" Stifel's note is the most explicit about valuation ceiling — INTC at ~$485B market cap, trading near TSMC parity on forward 2-year P/S. That math only works if you believe the foundry story is fully derisked. We're not there yet.
NVDA — Jensen made it crystal clear: this is a supply problem, not a demand problem. HBM, land, power, construction workers are the binding constraints. Demand doubles annually, supply can't keep up — industry can grow at most 2x per year. The Vera Rubin systems valuation at a trillion dollars is an order-of-magnitude larger than any prior generation. That's structural silicon consumption, not aspirational. The $500B+ LOI with SK Group locks in HBM4 for years. Bears calling a peak in the memory cycle are ignoring contractual frameworks that run through 2030.
MU — The Micron/Meta white paper quantified it: 38x slowdown when DRAM spills to SSD. Memory is a goodput enabler, not cost inflation. DRAM orders now stretch through 2027 as PC vendors lock CXMT supply. 64GB DDR5 server module contract prices ROSE 146% since end of June. Rate of change is accelerating. The market is missing Q in Revenue = P x Q — volume expansion in a supportive pricing environment drives revenues far above models. FCF generation sufficient to buy back entire float in ~5 years. Bears are betting on near-term catastrophe.
CXMT — IPO imminent at $85B market cap. Nomura price target implies $1T — a 13.4x divergence. Either the IPO price is deeply undervaluing Chinese DRAM strategic value, or Nomura is aggressive. Cost per bit remains 30% above leading suppliers, DDR5 die 40% larger than Samsung's. But Chinese pricing approaching parity despite yield gaps. PC vendors locking orders through 2027. Capacity expansions in Hefei, Shanghai, Beijing targeted for 2028 — massive ramp, execution risk high.
SKHY — Sold out through 2026, lead times stretching into 2027. Can't take more orders. The $500B+ LOI with Nvidia locks HBM4 supply for Vera Rubin. HBM consumes 3x wafer space per GB vs standard DRAM — every AI bit produced deletes three consumer bits. Structural tightening in overall DRAM market is hidden by headline supply numbers.
GEV — 116 GW gas turbine backlog, deliveries now pushed to 2031. Power/energy is the most supply-constrained node in the AI buildout, more so than chips. Market still pricing GEV as cyclical utility. Structural AI demand is permanent. Multi-year rerating opportunity if execution holds.
ADI — 73% gross margins defensible via design-in moat — customers rarely requalify analog components. AI opening new growth vectors for analog. But China domestic competition is the structural risk that deserves forward attention. New platform wins in China increasingly contested. Succession also a compounding governance risk.
META — World Cup tailwind but tough comps: 27% growth in Q3:25 vs 22% in Q2:25. Bar is high for Q3:26 guide. Public cloud/LLM API launch could embolden capex but opens new revenue stream. Low teens PE provides cushion vs S&P at 22x. $12B project-level financing for El Paso DC — new funding lever without diluting equity.
AAPL — Indirect beneficiary of others' AI capex without the capital burden. But CQ3/CQ4 estimates too high — rising semiconductor prices compress margins. 37x PE is expensive, pricing in perfection. Foldable phone with AI Siri is the long-term cycle catalyst, likely 12-18 months out. Smart glasses targeted for WWDC 2027 — privacy debate critical to success.
TER — UltraFLEXplus shipments more than doubled over 9 months, maintaining 12-16 week lead times. Test capacity is a critical bottleneck for AI chips. Photon 100 for silicon photonics positions TER as early leader in $300M-$700M market. Test times surging: Blackwell final test ~4x Hopper, Rubin 7x. Testing as % of GPU cost rising — secular tailwind.
AEHR — Effective backlog of $100.6M covers 67-77% of FY2027 revenue guidance. Exceptional near-term visibility. Concentration risk on few large customers is the key caveat — orders push out = volatility.
QCOM — Double-digit price hikes for smartphone processors pass through upstream cost increases. Protects margins, risks volume elasticity. Snapdragon Wear Elite driving edge AI smartwatch adoption — Q1 2026 shipments +70% YoY, penetration breaking 25%. New growth vector beyond phones.
AVGO — Samsung MOU valued >$200B through 2030 for memory, 2nm foundry, advanced packaging. Diversifies supply chain, strengthens custom silicon position. AMD Helios rack requires more than 550 Broadcom retimers — networking/interconnect products critical for next-gen AI.
KIOXIA — Leveraged ETF coming to US market. Mechanical rebalancing flows will amplify daily volatility. Added noise for long holders, setup for vol sellers. Flow mechanics distort price action, not fundamentals.
RIOT — AMD-Anthopic deal read as potentially increasing AI appeal of Rockdale facility. Thesis is narrative-driven and weak. No actual AI co-location contract. Crypto mining infrastructure not easily convertible to AI training.
GOOGL — Cloud revenue accelerated from +63% y/y to 82% in CQ2. Contracted backlog surged from $100B to $514B in 12 months — 5x increase. But raised 2026 capex guidance and reported first negative FCF quarter. ROIC debate is now concrete. Nearly 500 customers processed over 1T tokens — real usage data supporting growth.
AMZN — AWS reports this week. AI revenue disclosure granularity is the core focus. Bar is high after Google's 82% growth. X Energy SMR deal signals long-term power strategy. Shut down AGI Lab — focus on applied AI, positive for near-term margins.
TSM — First 2nm fab hit 20K wafers-per-month milestone. CoWoS capacity expanding from ~200K pcs quarterly to nearly 600K by late 2027 — 3x increase. Most important supply-side catalyst for AI chips. Arizona Fab 21 produced first GB300 chip. Doesn't want High-NA EUV yet — cost-conscious, likely pushes adoption to 2028.
SNDK — Same structural memory thesis: LTAs and SCAs lock in pricing floors while volumes expand. Market missing Q in Revenue = P x Q. Chinese hedge funds heavily long — smart money alignment, but crowding increases reversal risk.
Hearing AMD is rolling out optical interconnects (CPO) for MI500 in 2027. Channel checks point to Ayar Labs as the likely supplier. Optical interconnect inflection validates the networking layer thesis — keep an eye on CPO ecosystem names.
Word is Intel and SK Hynix Ohio JV confirmed by Ohio Senator. Structure still unclear — could involve HBM supply agreements, foundry partnership, or both. Multi-scenario setup. Gives SK Hynix US manufacturing foothold.
Channel checks suggest the GB300 chip from TSMC Arizona is real, but only CoWoS packaging remains to be localized. That's the next bottleneck to watch — CoWoS localization timeline. Market underpricing this gating factor.
Hearing CXMT IPO pricing at $85B vs Nomura $1T target is creating massive divergence. Some funds viewing the IPO price as deep value. Others see Nomura as aggressive. Either way, the gap signals extreme sentiment bifurcation — potential volatility on day one.
Word is META's Q3:26 comps are the real issue — 27% growth in Q3:25 creates a high bar. If ad revenue decelerates and capex guides up simultaneously, that's the most bearish combo. Cloud/API launch could embolden more spend. Burns two ways.
Channel checks suggest AAPL CQ3/CQ4 estimates are too high on both revenue and margin lines. Rising semiconductor costs are the culprit — memory and logic prices compressing margins. At 37x PE, any estimate miss could be ugly.
Hearing the AMD-Anthropic deal read-through to RIOT is narrative noise. Rockdale facility is not AI-ready without massive retrofitting. No actual co-location contract exists. Pure retail momentum trade.
Word is TSMC doesn't want High-NA EUV yet despite 98-second wafer times. Sees current EUV fleet sufficient. Pushes High-NA adoption to 2028 at earliest. Negative read for ASML's highest-value tool cycle.
Channel checks suggest Chinese hedge funds are heavily long Samsung, Micron, SanDisk, SK Hynix. Smart money alignment with memory cycle is clear. But crowded trades can reverse fast — positioning risk is real, not structural.
Hearing the Kioxia leveraged ETF coming to US will create mechanical rebalancing flows. For long holders, added volatility. For vol sellers, potential setup if fundamentals hold. Flow mechanics, not fundamentals, will drive near-term price action.