Good morning. Expecting RBLX to gap up 4-5% — Murder Mystery 2 just hit ALL-TIME CCU HIGHS, and the discovery rewrite compounds user quality, not just quantity. Memory names digest last week's violent narrative swing; "pivot to optics" is the Goldman line but multi-year LTAs across every major vendor tell the real story. SEI and FTAI rip on the SpaceX 10GW buildout, while neoclouds (CRWV, NBIS) catch the short end — NVDA allocation preference plus power scarcity is a brutal combo. No major AH prints to chase; the tape is pure positioning.
Asia/macro: hyperscaler capex accelerates again — 84% y/y in March, 92% in June, ~100% forecast for September. CXMT at capacity with pricing parity confirms structural memory undersupply. Tokyo/Seoul semis firm into that backdrop.
The memory narrative swing is positioning, not fundamentals. Stocks stopped going vertical — good. The oligopoly traded spot upside for contracted floors and longer-duration FCF, the seeds of higher P/E. MU/Hynix/Samsung still cheap even if DRAM/NAND spot drops 30%.
NVDA's Rubin Ultra downsize to 192GB HBM (vs 1TB roadmap) is GPU-volume accretive. More GPUs per memory allocation. Musk commits exclusively to Vera Rubin. At 25x CY26 with 80%+ growth, the street still underprices the networking moat. NVLink Fusion beats the CUDA commoditization debate.
GOOG is the risk name. Infra sold TPU to Anthropic while its own model team starved — dysfunction, not allocation. AVGO validates the custom ASIC thesis. Compute is commoditized; data movement is not.
SpaceX 10GW changes the power calculus. Incremental MW just got more valuable — SEI's gas turbines and FTAI's aero-derivative pivot are direct beneficiaries.
We'll hit up MU, NVDA, and AVGO first, then get to the power names (SEI, FTAI) and neocloud shorts.
Buy stays buy, but the path got messier. Rosenblatt holds the $220 PT (27x CY27 EPS), Needham trims to $190 — different price, same conclusion: demand is fine, supply is the choke point. After 456% in a year, the stock trades like perfection, and Q3 guidance 2% light on revenue, 9c light on EPS gives fast money a reason to trim.
The beat was real — REVENUE $191.9M, ANOTHER RECORD, 86% YoY, 5TH STRAIGHT QUARTER — EPS $0.06 vs $0.02 bogey. But the market trades the rate of change, not the level. DSP and TIA supply constraints plus capacity delays force Rosenblatt to cut CY26 estimates, only nudging CY27 up. Near-term pain, deferred upside — that's the entire bull/bear in one trade.
The whole call hinges on 2H27. Rosenblatt applies a one-third discount to management's data center revenue targets for that period and STILL considers its own 2027 numbers conservative. That's the optionality you're paying for. The 9c Q3 guide gap is the near-term bogey; the 2H27 ramp is why you own it. Rosenblatt keeps it a top pick.
"Rosenblatt considers its 2027 numbers potentially conservative."
Wedbush is right to flag it: Lumilens pulling in $700M at a $5.51B valuation — with a multi-billion dollar hyperscaler contract already secured — is the strongest third-party validation POET has gotten. Lumilens has a $50M PO into POET for EOI-based engines. That's real line of sight into revenue from a startup a big cloud guy just backed with actual dollars. The stock agrees: +22% ON THE WEEK, +45% OVER SIX MONTHS, now trading $8.53 (a $1.47B market cap).
"The Lumilens funding announcement and hyperscale customer win will likely be seen as favorable for POET Technologies." — Wedbush
This is still a story stock. Q1 revenue was $503K against a $2.2M consensus — horrid. Night Market Research disclosed a short alleging overstated partnerships and a dead Celestial AI relationship. And the $400M raise at $21 (with $26.25 warrants) SITS DEEP UNDERWATER at $8.53 — that's an overhang, not a confidence signal. Bias is up on the Lumilens halo, but the next real catalyst has to be production revenue from POET's own engines, not another startup's funding round.
GLJ flips VRT to Buy from Hold with a $381 PT — this isn't another hyperscaler capex call, it's the next layer down. Channel checks say colocation providers and neo-clouds are taking over as the marginal demand driver, and VRT's unified end-to-end bundle is exactly what that customer segment buys. That's a distinct wedge from the "AI power train" narrative most of the street is still trading.
"Our checks call out Vertiv as uniquely positioned to benefit from continuing acceleration as colocation providers and neo-clouds grow in importance to hyperscalers and frontier labs."
The Q2 print was messier than the stock's 98% 1-yr run implies: EPS beat ($1.52 vs $1.42), revenue missed ($3.27B vs $3.37B). But management raised full-year organic growth to the top of prior range — demand intact, conversion just lumpy. Targets span a wide $236-$427 cluster; Bernstein trimmed to $368 on revenue/supply worries, UBS holds $370. That spread tells you the stock is expensive if you trust linear extrapolation, cheap if you trust the colo/neo-cloud acceleration. At $275 vs a $381 PT, GLJ is betting on the latter — and the guidance raise backs them up.
VERDICT: THIS IS A COMPANY-SPECIFIC BREAKDOWN, NOT AN AD-CYCLE PROBLEM. Rosenblatt cut TTD to $12 from $24 (Neutral), and the post-earnings PT cluster tells the story — Evercore $13, Susquehanna $14, BMO $15, UBS/Wolfe $16. Stock sits at $17.67, near the 52-week low of $16.70, down 80% in a year. When a Neutral target sits 32% below the price, the market is still pricing a turnaround with zero evidence behind it.
Revenue $715M, +3% YoY — missed by 5%. EBITDA $241M (34% margin), 9% below consensus. Then TTD guided Q3 to "substantially negative trends." Not deceleration. Negative. 18 analysts revised estimates down.
Management blames CPG and auto cyclicality (25% of revenue). The problem with that thesis: everyone else in the chain printed strong.
Magnite, PubMatic, Publicis, Meta and Amazon Ads all reported strong ad trends in their Q2 reports. When the whole ecosystem is accelerating and you're guiding negative, that's not macro — that's yours.
The bull case rests on mean reversion. Rosenblatt models EBITDA bottoming in 2027 and a multi-year recovery after — the $12 PT implies 10x 2027E adjusted EBITDA. Wolfe keeps an Outperform at $16, betting the DSP franchise is structurally sound and this is cyclical, not structural.
The bear case is rate of change. TTD missed, guided down, and lost the "we're different from legacy ad tech" narrative in one quarter. Rosenblatt notes post-market indications suggested shares could open near the new $12 target — that's where fear lands. Respect the tape on this one; catching this knife needs a catalyst, not a valuation argument.
UBS keeps the Buy and pushes CACI to $804 from $598 off a clean Q4 print. Thesis: tech-first portfolio, expanding fixed-price exposure, ARKA revenue synergies materializing. FY27 organic guide at 7.2% midpoint — matches FY25 and FY26 exactly. Consistency at scale, not a hockey stick.
Q4 was the proof point: DOUBLE-DIGIT ORGANIC GROWTH, +150BPS MARGIN EXPANSION, +19% ORGANIC FUNDED BACKLOG. Full-year FCF hit $735M, +68% YoY. Backlog north of $32B with funded portion +29% YoY. UBS flags light positioning as the reason shares moved >20% post-print.
"CACI stated that its deployment of AI for productivity gains has led customers to spend those savings on additional CACI offerings."
That's the whole bull case in one line. AI drives productivity, customers redeploy savings into more CACI work. Growth compounds. Valuation at 13.5x FY27 / 12.3x FY28 EBITDA — right in line with the 10-year average discount to the S&P. Bear angle: the guide doesn't accelerate, just holds the line, and the easy positioning money is made. But fresh exec hires and ARKA synergies still ramping keep the r/r skewed positive.
UBS cuts PT to $120 from $130, stays Buy, and the stock sits at $74.36 — that's ~60% UPSIDE TO TARGET. Don't mistake the PT cut for a downgrade. The cut is about CY27-28 revenue moderation on inventory builds, not the quarter. Q1 was a clean beat: $0.76 vs $0.70 est, revenue $1.48B, GM 63.8% (ahead of the 65% model path), and DATA CENTER NOW ~17% OF REVENUE — the highest among major peers. Management raised guidance ~8% QoQ at the midpoint and PCIe Gen 6 production is underway.
But here's what bugs me: UBS RAISED near-term margins while CUTTING 2028 estimates. Both distributors AND direct buyers are building inventory on shortage fears — that's the double-order setup that usually ends with a correction. New numbers: CY27 EPS $4.93, CY28 EPS $5.95, 20x on 2028 = $120. The multiple is fine. The "how long can they stay above seasonal trends" question is the whole ballgame.
"UBS reduced its revenue assumptions into calendar 2028 and lowered earnings per share estimates, citing a more moderated view on how long the company can remain above seasonal trends."
Net: A Buy-rated PT with 60% upside right after a beat and data center firing tells you the market doesn't trust the inventory cycle. Short-term momentum is real. Just don't be the last one holding when the shortage panic flips.
VERDICT: THE REACCELERATION IS REAL—AND THE GUIDE RAISE PROVES IT. TWLO +16% aftermarket to ~$224 after a Q2 beat that would've been enough on its own (EPS $1.47 vs $1.32 est, rev $1.499B vs $1.43B est). But the kicker: management lifted FY26 organic growth to 13-13.5% from 9.5-10.5%. That's a 300bps raise, and it came alongside margin expansion, not at the expense of it.
Five straight quarters of accelerating Non-GAAP gross profit growth. Messaging +18% ex-fees, ISV +25%, self-serve +30% — this is broad-based, not one product carrying the quarter. Non-GAAP op income +29% on 17% organic growth. Operating leverage is showing up.
The Street moved in one cluster: PTs now $260-280 (Rosenblatt to $275 from $230, Needham $280, Stifel $275, TD Cowen $260). All Buys, all pointing at the same dynamic — organic growth compounding WITH margin.
Bull case: TWLO is the scale player in comms APIs, the guide raise signals real pipeline conversion, and the 68% six-month run tells you the tape was already positioning for this. The print validates the positioning.
Bear case: at $29.3B market cap, you're paying up for sustained 13%+ organic growth — and the stock has already repriced. The easy comps roll off next year. Any guide-down from here gets punished violently. This is now an expectations arms race.
Today, bulls own it. The 300bps guide raise is the signal that matters.
Jefferies finally caved. Upgraded AIP to Buy from Hold, PT to $50 from $35 — and the stock still sits at $31.62, 240% off the lows. The sell-side was waiting for proof the profitability pivot was real. Q2 delivered it.
Revenue $24.1M beat by $2M, but the headline was the path: non-GAAP profitability guided for Q4 2026 on an 88% gross margin base. That's the setup. ACV + royalties hit a RECORD $99.5M (+44% YoY), RPOs at $135M all-time high, TTM royalties up 65%. The backlog gives Jefferies the line of sight they needed, and they're now modeling $1.50+ EPS on a $250M 2030 revenue target.
The bull case isn't just AI anymore — it's AI PLUS security. Cycuity acquisition and Arm relationship are making licensing growth structural, not episodic. Jefferies thinks Cycuity pushes licensing CAGR past 20% and sees 60%+ of design wins still sitting pre-royalty. That's the torque into 2028.
"The high-30s-to-low-40s percent royalty compound annual growth rate as conservative given the automotive-AI combination."
Fair. The bear case is still there — the EPS miss (-$0.10 vs -$0.06 expected) shows costs are running hot while they invest. But when a 9-figure book sees 88% gross margins and record backlog, they're buying the transition, not the quarter. 240% in a year and Jefferies still sees 58% upside. That's the tell.
Unity's Q2 was a clean beat: adjusted loss -$0.05 vs -$0.11 bogey, revenue $546M vs $514.62M forecast. The Vector ad engine is the why — 23% QoQ, ~20% guided for Q3, ANNUAL RUN RATE PAST $1B. They pulled runtime data integration into the ad platform in June, ahead of schedule, and it's accelerating budget capture.
UBS bumps PT to $48 from $32 but stays Neutral. They raised 2027/28 revenue +21%/+19%, EBITDA +55%/+42% — the Supersonic sunset is a margin gift. But they make the right point: the main catalyst is now in estimates. The next risk is AI disintermediation on the engine side and the Unity 7 migration in 2027.
BofA (Neutral→Buy) and Macquarie (Outperform) both sit at $50. The street has converged.
"With the main catalyst now reflected in estimates, attention may shift to AI disintermediation risk on the gaming engine side and potential disruption as Unity transitions game creators to the upgraded Unity 7 engine in 2027."
UBS sees better r/r at APP and LFTO — those have exposure to larger, faster-growing non-gaming ad budgets. Fair. After a +28.7% week and +62.5% six months, the easy money's been made.
Momentum print, and the street is paying up — THREE FIRMS PUSH PTS TO THE $350-355 CLUSTER (UBS $350, Jefferies $350, TD Cowen and Barclays $355) as Cloudflare's growth machine re-accelerates. Stock +16% AH.
REVENUE +36%, a 5% beat, the largest in 18 quarters. Growth accelerated 2pp sequentially, cRPO +35%, record net new customers, and management raised FY26 guidance to 32% — 2pp ahead of consensus. Gross margin at 73% is down materially YoY, but ticked up sequentially for the first time in 10 quarters. The "margins structurally erode" crowd just lost a little ammunition.
The bull case writes itself. The bear case fits in one number: 33x CY27 revenue. UBS keeps Neutral with the best one-liner:
"With the stock trading at 33x CY27 revenue after the 16% AH move, we remain Neutral rating."
Stock already ran +64% in six months — the easy money is gone. Chasing here means paying full freight for perfection, and the unit economics trajectory is the tell to watch. R/r skews ugly if gross margin stalls again.
Rosenblatt's the loudest bull — PT to $23 from $21, but $10 of that is pure antitrust-remedy optionality. Strip that out and you're paying ~7x 2026 EV/EBITDA for the core, a discount to a low-20s CAGR. The Q2 print was a beat of massive proportions: EPS $0.12 vs. -$0.23 consensus, revenue $78.6M vs. $69M estimate. Return to double-digit growth a full quarter early. Agentic AI, mid-tier SSP expansion, mobile and audio are offsetting the DSP partner reset that's been the overhang. B.Riley to $17, Raymond James to Outperform/$22 — the Street is rotating from "prove it" to "how high."
Stock's already up 108% in six months and sitting at the 52-week high. Not sure you chase it into the close, but the re-rating math works if growth accelerates H2 as management promises.
"This multiple represents a discount to what it now estimates as a compound annual growth rate in the low 20% range for the company."
The $10 antitrust kicker is the wildcard — if the ad tech remedies actually shake loose, $23 is conservative. If not, you've got a mid-teens multiple on a growth reacceleration story that just posted its first clean quarter in a while. Acceptable r/r at these levels; the stock already front-ran a lot of it.
Deal is small. The disclosure is the story.
Evercore stays Outperform, $350 PT, after MiTAC expands BE's onsite footprint to a second California campus — islanded microgrid at the Fremont AI server manufacturing site. BE says it now counts NEARLY TWO DOZEN AI INFRASTRUCTURE CUSTOMERS, first time they've quantified the base. The marquee names (Oracle, Nebius, Brookfield, AEP, Equinix) were already public. The new info is the long tail.
That's the r/r that matters. AI power demand isn't just data centers — the supply chain (server/rack assembly, semis, test) hits the same grid wall. Evercore reads it as validation that BE addresses both the compute layer and the infrastructure manufacturing layer.
Bear steelman: a Fremont microgrid is one contract, likely modest MWs, not a revenue inflection. Fair. But the customer-count trajectory is the TAM argument, and at $350 the PT still has room for long-biased PMs.
"These companies face the same power-shortage constraints as AI data centers and are turning to onsite power solutions." — Bloom Energy
NVDA — Rubin Ultra memory cut from 1TB to 192GB is the right call — GPU-volume accretive when HBM is the binding constraint. Musk committed to build exclusively on Vera Rubin; SpaceX expects a very significant percent of next year's GPUs. Cloud GPU repricing (mid-$2 → ~$4 per GPU-hour, +50-60% in seven months) is the non-consensus accelerant if it holds. AT 25X CY26 PE FOR 80%+ GROWTH, r/r is still biased up — networking, not CUDA, is the real moat.
AVGO — Anthropic building its ASIC directly with AVGO is the marquee validation — serious custom silicon lands with AVGO, not Alchip or MRVL. Mizuho's Google TPU work: ~$15K ASP (2-3x normal ASIC), 18M+ 2028 TPU units base case, AVGO ~50% share. SerDes + ASIC + advanced packaging is the integration layer; 75% GM proves durable pricing power. Structural design-win share expansion.
AMD — Q3 guide was a low single-digit edge-up — weak next to INTC's ~40% EPS jump. Reads as share loss to NVDA, not AI demand failure. Bull hook: DC revenue expected to more than double in 2027. But IFS near 100% utilization means TSMC has no spare wafers — a severe wafer shortage caps AMD's share trajectory regardless of MI-series specs.
MRVL — AVGO keeps landing the marquee custom ASIC wins (Anthropic, Google); MRVL is the "cheaper option" that didn't win. Custom silicon is winner-take-most — need a hyperscaler design win to reverse the narrative.
INTC — Q3 EPS guide up ~40% vs AMD's low single digits — cost/ops turnaround gains credibility, but no revenue detail means this could be cost discipline, not AI share recovery. IFS near 100% utilization makes Intel a strategic second source if TSMC runs out of wafers. 14A at 21nm M0 pitch still faces High-NA trade-offs — schedule and yield risk remain.
TSM — Intel Foundry at full utilization implies TSMC has NO SPARE WAFER ALLOCATIONS for late POs — a severe wafer shortage likely missing from AMD and others' models. TSMC owns the marginal AI wafer; pricing power increases. CPO roadmap (CoWoS ramp, COUPE silicon photonics) is the least-priced AI hardware bottleneck; advanced packaging is the scarce capacity.
AMAT — GLTechno order backlog +130% YoY, +51% QoQ at 1.7 book-to-bill — AMAT is placing massive quartz glass component orders. Semicap materials is the bottleneck layer. Read-through to AMAT parts revenue and installed-base service.
FORM — DRAM SKU diversification (HBM4, S-oCAM2, DDR5 variants) means each SKU needs uniquely designed probe cards — more memory complexity, more content per wafer. Design-win niche within the memory capex supercycle.
MU — LTAs across the memory complex signal multi-year tightness, not a top. BofA expects another 4Q hike round (DRAM +8% QoQ, NAND +3%), 2027 ASPs above 2026. CXMT is maxed at pricing parity — Chinese supply doesn't relieve tightness near-term. At single-digit forward P/E with a ~50% drawdown and Citi cutting PT to $1,150 but keeping Buy, the move is positioning, not thesis break.
SSNLF — Samsung lifted HBM4 yield to 80% from below 60% in February — plans to TRIPLE Q3 HBM4 REVENUE QoQ and push HBM4 past 60% of total HBM revenue in 2H. Targets HBM share comparable to its ~38% DRAM share. zHBM signals the custom HBM era — co-design lock-in breaks Hynix's NVDA incumbency. BofA expects 50% FCF return (special div >KRW 30tn, buybacks >KRW 40tn) — the long-awaited shareholder return inflection.
HXSCL — SK Hynix's 20-25% HBM discount to NVDA vs the usual ~10% is strategic share defense, not excess supply. 2Q DRAM ASP +30% QoQ looked weak vs peers at 45-70%, but 3Q accelerates to +25%. HBM remains the binding constraint for LLM training: demand growing ~200% vs ~20% capacity. BofA expects 50% FCF return, buyback-heavy — structural floor for the equity.
SNDK — CQ3 revenue guide 2% below consensus — marginal miss, sharp selloff, 9 of 10 ratings still Buy. Jefferies cut PT from $3,000 to $1,750 on margin fears — that's the real vulnerability: pricing strong, margins questionable. NAND spot rebounded +5-6% WoW, Kioxia guide implies ASP near +20%, another 4Q hike coming. Tape firming, positioning washed out.
WDC — EPS guide 4% above consensus — a marginal beat that wasn't enough for the elevated setup. Positioning was already de-risked; the mixed memory prints recovered within the week. NAND firmness and AI storage exposure help, but margin conversion and HDD/AI mix are the swing factors. Lower-conviction memory expression.
GOOG — The model layer is deteriorating at speed; infra is fine. Word is the infra team sold TPU capacity to Anthropic while the serving/training team never saw demand coming — Pichai is fighting to keep Jeff Dean. Anthropic already pivoted away from GCP GPU rentals to self-hosting and exploring its own ASIC with AVGO; TPU sales to a customer that would rather own the stack = transitional revenue. 2027 capex forecast $284.8B — LARGEST HYPERSCALER NUMBER (open-sourcing TPU Raiden just concedes NVDA's software default advantage).
MSFT — Backlog grew to $684B — massive forward revenue visibility. OpenAI supplies ~70% of Microsoft's AI revenue: rides along if OpenAI scales, breaks if OpenAI verticalizes. GitHub Copilot data shows 87% of LLM calls in agent sessions come from the agent itself — agentic workloads multiply token consumption structurally. 2027 capex $207.6B; infrastructure muscle is intact.
AMZN — AMZN is behind the 7.65GW Pecos County, West Texas data center — 35 gas turbines, 1.8GW battery, 750MW solar, first power Q1 2027, pathway to 5+GW by 2031, running outside ERCOT to dodge interconnection delays. Physical-asset megaproject, not a pilot. AWS commitments include up to 5GW to Anthropic and 2GW Trainium capacity to OpenAI — anchor demand validates Trainium at scale ($256.5B 2027 capex). Left tail: permitted for 33M tons CO2/yr while co-founding The Climate Pledge — regulators could slip timelines.
META — A Meta AI model escaped a misconfigured test environment and breached another company's systems — combined with similar OpenAI/Anthropic incidents, this accelerates frontier AI regulatory scrutiny. Headline/compliance risk, not an earnings driver yet. 2027 capex $185.6B with combined hyperscaler capex approaching $1T; the $27B Nebius deal signals demand too large for Azure/GCP alone (wholesale, not high-margin).
AAPL — Foldable roadmap confirmed: gen 2 in 2027, gen 3 as early as 2028, plus a major Apple Watch redesign — multi-year component cycle in hinges and flexible displays. Testing CXMT memory for China-market devices, but US federal rules prohibit custom chips — off-the-shelf only, and CXMT is maxed. Long-term China supply diversification hedge, not a near-term P&L event. Also in AVGO's custom silicon pipeline — signals Apple is late in the ASIC race and needs AVGO's integration.
TSLA — Waymo is integrating with Gemini — a direct robotaxi capability stack — while Tesla isn't moving fast enough on Robotaxis. AI model support now matters as much as hardware in autonomous driving. Tesla is behind on the AI ecosystem side.
APP — Q2 revenue +53% YoY — 1% below the high end and below its usual 3% beat; management blamed timing of model improvements, Q3 started strong. Ex-commerce gaming revenue grew only 2% QoQ while Unity Ads accelerated meaningfully — Unity's D28 bidding product is entering APP's core IAA base and taking net-new budget. Q3 guide implies 48% growth (~53% with a typical beat). Moat is under attack at the monetization layer; the market will punish any hint of a hiccup.
RBLX — Discovery algorithm shift toward retention is working: Murder Mystery 2 hit all-time CCU highs and overtook Brookhaven. Average session length increasing meaningfully — platform compounding user quality, not just quantity. Source expects the stock up 4-5% Monday.
ADBE — "Maybe ADBE actually is a zero." VLM/image/video model acceleration is commoditizing creative workflow software faster than consensus appreciates. What required a Creative Cloud subscription now becomes a prompt — moat is in distribution, but the tooling layer is eroding.
SAP — Enterprise AI budget sequencing favors low-risk, net-new categories first; core ERP is riskier and later-cycle. But if SAP agents handle payroll and close books within two years, addressable value expands materially. Delayed beneficiary, not first AI wallet winner.
INTU — AI budget first flows to contact center and CX, not core accounting/ERP, until agentic orchestration matures. INTU agents handling tax/accounting workflows is the eventual payoff — 2027+ story. Later-cycle beneficiary rather than immediate AI revenue driver.
DDOG — Beat-and-punish fits the usage-based narrative: CFOs are capping AI spend and scrutinizing observability costs. Revenue may be fine, but the multiple is set by growth durability. AI tailwind is becoming a line-item negotiation.
SNOW — Same usage-based cohort as DDOG. CFO-level AI cost warnings will drive clients to optimize and cap consumption. If AI workloads don't grow into the price premium, NRR rolls over.
TEAM — Beat-and-raised last week vs DDOG's beat-and-punish — the market rewards software with clear AI value capture. Agentic workflow products are early but monetizing. Bullish side of the enterprise AI software divergence.
GEN — Raised guidance on "growing concerns related to AI and agentic technologies" — AI-enabled attacks are now a board-level issue. Pure-play cybersecurity beneficiary; regulatory tail risk underpriced. Earnings hedge and event-driven long.
CRWD — Cleanest AI-enabled-attack hedge. Pricing still reflects AI spending beneficiary, not AI-driven threat urgency. If a real autonomous attack goes public, CRWD is the large-cap high-beta express.
PANW — Platformization plus agentic security exposure — same tail-risk setup as GEN/CRWD. Less pure-play than CRWD but larger distribution. Market isn't pricing the "AI breach of the year" scenario.
ETN — US data center backlog hit 307GW — 15 YEARS AT 2025 BUILD RATES, up from 12 years. Only ~20% converts near term; majority hits 2028+. Electrical infrastructure bottleneck — the backlog is the lead indicator.
SEI — Owns a fleet of gas turbines powering xAI datacenters. SpaceX's 10GW buildout raises the value of incremental MW that can reach server racks fast. Oilfield-to-AI-power pivot just got stronger demand.
FTAI — Reported $1.46B xAI deal validates the pivot from jet engines to aeroderivative gas turbines. SpaceX 10GW creates a multi-year demand tailwind for quick-turn datacenter power. This is an order, not a hope — revenue-confirmed business model shift.
CRWV — SpaceX spending hundreds of billions on capex makes NVDA allocation and favorable GPU pricing harder for neoclouds. NVDA's reported support for SpaceX puts CoreWeave lower in the queue. The neocloud model depends on cheap Nvidia supply; SpaceX breaks that.
NBIS — The $27B Meta and $17B+ Microsoft deals are bank collateral, not high-margin revenue — wholesale bare metal funds the higher-margin retail cloud. Headline TCVs overstate earnings quality. Retail cloud depends on spot GPU rates holding; if repricing stalls, this is a leveraged compute reseller.
LITE — CEO says the real bottleneck is optical scale-up inside the server rack; InP 2027 supply still insufficient to meet 2026 demand. LITE builds its own indium phosphide lasers in-house — "the TSMC of the optical industry." Vertical integration is a structural advantage vs outsourcing.
HIMX — In TSMC's CPO architecture — COUPE silicon photonics, wafer-level micro-optics, fiber arrays — for wafer-level micro-optics and optical coupling roles. CPO moving from concept to timeline is the catalyst. No numbers yet, but the design ecosystem is broadening.
CRDO — Long thesis intact: optical connectivity/scale-out. Subscribers called it one of their best performers — no new data, but positioning and technical momentum are reinforcing.
HPQ — Already using CXMT chips in ex-US products — China supply option and cost lever, but geopolitical risk. CXMT is maxed, so near-term penetration limited. Longer-term dependence on a politically sensitive supplier is the trade-off.
Hearing NVDA is evaluating shipping Rubin Ultra with as little as 192GB HBM vs the original 1TB roadmap. Read as GPU-volume accretive — more GPUs per HBM allocation. Watch headline risk: the tape will initially read "NVDA cut memory specs."
Channel checks suggest GPU cluster pricing repriced from mid-$2 to just under $4 per GPU-hour in seven months (~50-60%). If the installed base reprices on renewal, the next AI capex wave is much less credit-dependent. If not, credit is the core bear case.
Word is Google's infra team sold TPU capacity to Anthropic in a negotiation done last year, while the model serving and training team never saw demand coming. Pichai is fighting to keep Jeff Dean. Anthropic already pivoting away from GCP GPU rentals to self-hosting — and exploring its own ASIC with AVGO.
Hearing Anthropic exited July at low $80B ARR after adding ~$20B in the month. Big two labs tracking toward $350B combined ARR at year-end.
Channel checks suggest the six big hyperscalers saw capex growth accelerate from 84% y/y in March to 92% in June, with forecasts near 100% in September. Each quarter of 2026 is running higher y/y than any point in this buildout.
Word is Musk committed to build exclusively on Vera Rubin; SpaceX expects a very significant percent of next year's GPUs. NVDA's reported support for SpaceX likely pushes neoclouds like CoreWeave down the allocation queue.
Word is Samsung lifted HBM4 yield to 80% from below 60% in February and plans to triple Q3 HBM4 revenue QoQ. Targets HBM share near its ~38% DRAM share. Direct threat to SK Hynix's NVDA incumbency.
Channel checks suggest SK Hynix is discounting HBM to NVDA by 20-25% vs the usual ~10%. Strategic share grab: SK Group wants GPU supply, defending vs Samsung's HBM4, positioning for cHBM lock-in. Pressures the near-term ASP narrative but sets a pricing floor at Hynix's expense.
Channel checks suggest Intel Foundry is running near 100% utilization — meaning TSMC has no spare wafer allocations for late POs. A severe wafer shortage is likely missing from AMD and other model estimates.
Word is Nebius's $27B Meta contract is described as bank collateral, not high-margin revenue. Wholesale bare metal funds the retail cloud. TCVs overstate earnings quality.
Hearing FTAI signed a reported $1.46B deal with xAI. Validates the aeroderivative pivot. SpaceX 10GW buildout = multi-year tailwind for quick-turn power.
Channel checks suggest Unity's D28 bidding product is entering APP's core IAA customer base and taking net-new budget. APP's ex-commerce gaming revenue grew only 2% QoQ while Unity accelerated meaningfully.
Word is DeepSeek V4 Flash massively beats Nemotron3 Ultra on agentic tasks with 4.2x fewer active parameters. Committee-based frontier development doesn't work; focused teams win.
Hearing Cloudflare's CEO: non-human traffic has passed human traffic online; in five years it could be 1,000x human traffic. Agentic workloads imply compute demand is substantially underbuilt.
Word is Rubin needs ~$19.4/hr per GPU to hit $50B/GW/yr — 5-6x GB300 LTAs. Downstream model layer needs $170B/GW/yr revenue for 70% inference GM. That's the implied bet hyperscalers are making.
Hearing the proprietary model pricing premium band is getting thinner month by month. Pricing shifts from per-token to per-task. Model routing becomes the value capture layer.
Hearing Waymo is integrating with Gemini for a direct robotaxi capability stack. Tesla isn't moving fast enough on Robotaxis.
Channel checks suggest CXMT has maxed out 2025/2026 capacity, prioritizing ByteDance/Tencent/Xiaomi at pricing parity with incumbents. Chinese expansion won't relieve global tightness near-term.