Wednesday, August 26, 2026

Good morning.

SOX keeps derating — 21–22x down to ~15x forward — and credit leads equity on AI risk. NVDA/AVGO CDS at record wides. Goldman's Delta-1 desk has the trifecta (credit, inventory, power). Story survives, multiple doesn't. The counter-narrative is the Jalapeño print: OpenAI's first-gen ASIC beat Blackwell/Rubin on perf/W and perf/$ — first credible challenge to the CUDA tax. AVGO the design partner, CLS the ODM; CLS the cleaner expression at under 16x 2027 GAAP EPS. Broadcom's even financing the NRE — Hock converting balance sheet into design-win lock-in. Memory's the other big thread: TRENDFORCE HAS DRAM+NAND AT 68% OF CSP CAPEX IN 2027, HBM CONTRACTS UP ANOTHER 70–140%. MU/WDC eat well there. SMTC printed record backlog and guides +45% q/q data center revenue — optics upcycle confirmed, read-through to LITE/COHR. Colo execution beating the bear case per MS — APLD/RIOT/CIFR/CORZ on time/on budget, WULF the cost slipper. Power remains the binding constraint; ERCOT congestion the live case study. META agent appstore in WhatsApp is the sleeper — WeChat mini-programs with 3B DAUs, zero consensus dollars modeling it.

We'll hit up AVGO, CLS, and MU first, then get to the colo/datacenter group.

CORE ANALYSIS

PLTR

THE FDE MOAT

Truist's back with a $223 PT, but this time the angle isn't the software — it's the humans. They hosted an AI recruiting firm president who spent 6 months on 250+ C-suite interviews and 300+ AI professionals, and the takeaway cuts straight to the core of Palantir's structural advantage: the scarcest asset in AI isn't compute, it's forward-deployed engineers, and PALANTIR TRAINED ~80% OF ALL ELITE FDES.

That's the moat that shows up in the P&L — 84.8% gross margins don't happen by accident. Anthropic and OpenAI are "talented but still early" trying to replicate the FDE model. And the recruiting guy's outlook for IT services is bleak, which makes sense: they're the legacy version of the same promise, and Palantir built the human layer that actually delivers it.

"Having pioneered the forward deployed engineer model, Palantir has trained approximately 80% of all elite forward deployed engineers."

The post-Q2 analyst cluster keeps building. 22 firms revised estimates upward after the print — revenue 6.8% above consensus, op income 10.5% above, FCF 9% above. UBS to $220, Piper at $230, Phillip at $215, Truist holding $223. Benchmark's the lone Hold, but even they're conceding the Rule of 155 (up from 127) — absurd for a company growing revenue 93% with U.S. Commercial reaccelerating to 149% YoY.

Stock at $175.89 vs. $230 high PT. The bull case is structural: nobody else has the FDE pipeline. The bear case is valuation, always. But at this rate of change, the bears are fighting the tape.

CRWD

EARNINGS IN TWO DAYS and the tape's mixed — one OW reiterated, one PT raise, one trim, one Neutral. But the actual signal this morning isn't the quarter, it's the CTO walking.

Cantor sticks at Overweight, $250 PT (split-adjusted, up from $181 basis). KeyBanc raises to $240 on strong partner checks. Scotiabank trims to $227 on valuation — still Sector Outperform. Guggenheim sits Neutral, saying the numeric setup matches guidance but "presents challenges." Net: nobody pounding the table, nobody running for the exits. STOCK AT $191.95, +64% YTD, already pulled back.

The real story: CTO Elia Zaitsev leaves after 13 years to co-found Cognition, an early-stage fund targeting agentic AI security threats, with two ex-CrowdStrike corp-dev guys (they previously founded Brightmind). 3-4 seed investments a year. Cantor reads the departure as conviction, not disruption:

Cantor Fitzgerald views the move as investor conviction that agentic AI could drive a platform reset across the security stack.

That's the thesis worth watching. Zaitsev's not retiring — he's betting that AI-driven attack surface expansion births a new security stack, one CrowdStrike doesn't own. If he's right, the moat gets tested from a new direction, not by another endpoint vendor. If he's wrong, it's a footnote and CRWD grinds higher on the core business. Earnings answer the latter; the fund tells you where smart insiders think the puck is going.

ESTC

TD COWEN JUST PUT A $70 PT ON A STOCK TRADING AT $83.24. That's the whole ballgame. They lifted from $55, but this is catch-up after the stock ripped 42% in a month — not fresh conviction.

The underlying thesis: modest beat to Sales-Led Subscription growth (guided 16%) with flow-through to an FY raise. The channel disagrees. Partner checks flag heavy competition from OpenSearch and security vendors, plus a total absence of AI demand. Columnar Mode helps retention, but TD Cowen reads it as proof the architecture needs a rethink — not a catalyst.

"A partner check conducted by TD Cowen revealed caution in the market, citing high competition from OpenSearch and security vendors as well as lack of AI interest."

Street's fully bifurcated. Stifel at $90 (Buy) and Piper at $85 (Overweight) ride the growth acceleration. Cantor's $91 is Neutral — a multiple expansion call, not fundamentals. Morgan Stanley cut to Equalweight/$66 on near-to-medium growth concerns. That's the bear camp: the 42% run has front-run the fundamentals.

At $83 you're paying for the bull case. Downside to TD Cowen's target: ~16%. Upside to Stifel's: ~8%. Unless SL Subscription inflects hard at the print, the r/r is ugly. The easy money is gone. (Watch the guide — the bar just went up, not down.)

NVT

UBS stays Buy with $225 PT after NVT drops $1.75B on Maverick Power. That's 11.5x 2026E EBITDA for incremental data center power distribution — cheap if AI capex keeps ripping. (About 6% EPS accretion to 2026, per UBS — not life-changing, but the right kind of bolt-on.)

UBS analyst Neal Burk stated the deal is valued at 11.5x 2026 anticipated EBITDA.

Q2 WAS THE REAL PRINT: EPS $1.45 vs $1.16 est, revenue $1.5B vs $1.26B, sales +53% YoY with 47% organic. Management raised FY26 guidance off data center and infrastructure demand. Evercore ISI bumps to $210 on the 23% EBITDA beat and execution against the $2.5B backlog. 13 analysts revising upward. (Stock's up 72% in a year — this isn't a discovery.)

Bull case writes itself: AI power distribution is the highest-conviction spend line in tech infrastructure, and NVT keeps buying niche capability at reasonable multiples. Bear case: 72% run, crowded trade, premium creep on strategic M&A. At 11.5x for this end-market, r/r still skews positive. Just don't chase size.

PANW

Cleanest setup into a print you'll see all week. JPM raises PT to $384 from $326 (Overweight) on platform momentum, but the broader sell-side cluster is even louder — BofA at $420, Cantor at $425, Benchmark at $400, UBS at $390. Everyone's pointing the same direction into FQ4 earnings, which land in SEVEN DAYS.

The thesis is simple: post-Mythos, CIO urgency around security posture has accelerated, and PANW's platformization strategy is the direct beneficiary. JPM's Brian Essex notes pipelines are accelerating across the coverage universe — that's a read-through for the whole security complex, not just this name. The "right place, right time" narrative is doing heavy lifting here, and honestly, it's earned.

Financials back the setup: $3.79B levered FCF LTM, revenue +19.5% to $10.6B, 72% gross margin. The swing factor is NGS ARR — JPM sees "potential upside" to the outlook. Cantor's leaning on the new Frontier AI Critical Defense Program as the next demand driver. Hardware revenue and NGS ARR growth are the two bogeys BofA flagged.

"The company is in the right place at the right time."

That's the whole ballgame. Beat on NGS ARR and this stock runs to the $400+ cluster. Miss, and the PT floor gives you cover. 7 days.

NVDA

NVDA beats tomorrow. BMO — the only fresh voice today — sees Q2 landing $2-3B above the $92B Street bogey and Q3 guiding another $2-3B above $104B, all on data center strength. The real signal is rate of change: sequential adds step from ~$4B/quarter in FY25 to $10-12B/quarter in 2H FY26 as Vera Rubin goes vertical.

"The VR NVL72 plus Groq3 LPX delivers 35 times higher inference throughput compared to GB200 NVL72, leading to higher revenues."

That step-function is the whole ballgame. Rest of the Street already lined up behind the same thesis — Raymond James at $352, Cantor at $350, KeyBanc at $330, Rosenblatt at $325 — Rubin R200 in mass production since July, CPU revenue creeping toward 5% of sales by 2028. Demand "remains robust." No serious bear case into this print. The only debate is whether 35x throughput gets one turn or two.

CRM

Truist says the cheap multiple is the story. Reiterates Buy at $280, anchoring on 12X CY27 FCF and a 0.61 PEG. That's the bull case in two numbers — and they're holding the line even with bears shouting about H2 ORGANIC GROWTH and agentic AI not being a real catalyst yet.

The quarter itself: Truist thinks Q2 strategic areas progressed solidly and sees improving organic growth ahead, driven by favorable net new annual order value trends. So the move is forward-looking — they're betting the growth deceleration troughs, not arguing with today's tape.

The bear case gets a fair shake: long-term model achievement questioned, acquisitions muddy the story, and the "value trap" label won't die. Fair enough — if agentic AI stays a narrative and organic growth doesn't inflect, cheap can get cheaper. Truist's retort: at 12X forward FCF, the bad news is already in the price.

"The firm maintained its positive stance despite ongoing concerns about the company's second-half organic growth inflection and questions surrounding agentic AI as a material catalyst."

Not a crowded desk today — one article, light tape. Peers split-ish: Citizens at $315, Oppenheimer constructive into the print, Cantor more conservative at $250. CRM reports F2Q27 next — that's the real catalyst, not the analyst chatter.

RBRK

BULLISH INTO THE PRINTS. Scotiabank bumps PT to $114 from $95 (Sector Outperform), and they're just one of five firms stacking the deck ahead of Thursday's F2Q report. Cantor to $120, Oppenheimer to $120, Guggenheim at $110, GS at $106 — the cluster is tightening around a beat-and-raise setup, and the tape reflects it: stock at $98, up ~50% in three months, $20B market cap.

The thesis is simple: Rubrik is the AI-preparedness trade in backup/recovery. Security-first architecture gets the budget before the LLM data lands, and Identity Resilience is surfacing as a priority in fieldwork. Street is feeding the flywheel too — EIGHTEEN ANALYSTS REVISED ESTIMATES UP into the print.

"Rubrik's positioning among backup and recovery vendors to benefit from AI preparedness spending due to its security-first architecture."

Two risks to carry into Thursday. First, the GTM bleed: ex-CRO Brian McCarthy is raiding Rubrik for Cursor, and ~30 senior/quota-carrying sales folks followed him out the door. That's a real overhang on execution, even if the quarter is clean. Second, elevated hardware pricing/lead times — still a margin and duration question nobody's answering yet.

THE VISIBLE SETUP

Thursday's guide matters more than the beat. Consensus expects a Q3 guide above Street and a fiscal 2027 revenue raise — that's already in the stock. If they deliver the raise AND hold gross margin, the next leg to $120 is easy. If hardware costs leak into the guide, we get a "good quarter, bad print" moment. Longs are positioned for perfection. r/r is fine, but don't chase into the print — let the guide tell you which Rubrik you own.

SNOW

THE SETUP GETS CLEANER

Cleanest setup in years. Benchmark pushed the PT to $360 from $290 into the Sept 2 FQ2 print and kept Buy. SNOW sits at $321.26, up 66% in a year, right under the $341.95 high. The firm sees product revenue beating the 30% consensus and op margin clearing the 12.5% bogey. Channel checks across hyperscalers and data platforms come back clean.

THE HYPERSCALER BACKDROP IS THE TELL. Oracle OCI +93%, Google Cloud +82%, Azure +43%, AWS +37%. Benchmark's top-3 average hits 54% — the healthiest since 2022. Enterprise AI demand is capacity-constrained, not interest-constrained. That's the wind at SNOW's back.

THE DEBATE MOVES TO GUIDANCE

Beating FQ2 is table stakes. The real question: can management lift FY27 product revenue guidance past 31% growth? Post-Q1 and June's Investor Day, expectations already run hot. Street splits on direction, not conviction — Morgan Stanley Overweight (34% growth), TD Cowen at $370, Stifel at $350, JMP upgraded. Even Guggenheim's Neutral carries no heat, admitting execution's solid.

"The key debate has shifted from whether Snowflake can outperform quarterly expectations to whether management has sufficient visibility to raise fiscal 2027 product revenue guidance beyond the current 31% growth outlook."

MRVL

THE SETUP INTO THURSDAY'S PRINT IS AS GOOD AS IT GETS FOR A $200B SEMI — AND THE BAR IS HIGH.

Susquehanna bumps PT to $265 from $230 (Positive) into fiscal Q2 earnings, and the thesis is simple: AI networking is compounding faster than the market prices, and the custom silicon story just got a whole lot bigger. The Google deal is the headline — roughly $120B OF CUSTOM PRODUCT BUSINESS THROUGH 2033 across storage controllers, NICs, and AI inference accelerators. That's not a partnership, that's a franchise.

The custom business math is where the hockey stick shows up: Marvell sees custom more than doubling in FY28 to ~$4B, then nearly tripling again to ~$10B in FY29. And Susquehanna expects the company to RAISE ITS FY27 TARGET FOR THE INPHI BUSINESS — that's the optical networking crown jewel, and the read-through from Coherent and Lumentum keeps getting louder.

"Susquehanna said read-throughs from transceiver module and component makers including Coherent and Lumentum continue to be robust."

Supporting firepower is thick: Rosenblatt at $300, UBS at $310 post-Google, Benchmark at $275, BMO and Oppenheimer at $250. The street has basically converged on the same narrative — custom ASICs plus optical interconnects is the AI arms dealer play, and MRVL has the only credible #2 position behind Broadcom.

Caveat: shares already up 215% over the past year and trading above InvestingPro fair value. That's the bull case priced in — but AWS raising capex to $220B (even if memory-driven) and the Google agreement give the fundamental spine. The risk isn't demand; it's whether Thursday's guide can clear what's now an extremely crowded long. Watch the Inphi raise — if it doesn't show up, this thing gives back the pop quickly.

1. Supplementary Coverage

AVGO

Jalapeño validates AVGO as the go-to merchant custom-silicon house beyond Google/TPU. Street math says the >$100B FY27 guide already banks the ~10GW program — headline trades as done. Marginal signal is credit: CDS +5bps to a record wide, bondholders done subsidizing AI capex. That's the swing factor, not another design win. (Thor Ultra 800G NIC: underappreciated anti-NVDA attach.)

CLS

Cleanest expression of the Jalapeño trade — ODM partner at <16x 2027 GAAP EPS that consensus hasn't marked. Production-ready by year-end, Gen 2 deep in development. If the ~10GW framework scales, CLS gets a multi-year ODM volume tailwind without AVGO's credit overhang.

MU

TRENDFORCE: DRAM+NAND HIT 68% OF CSP CAPEX IN 2027, HBM CONTRACT PRICES +70-140%. Strongest pricing-power setup in the AI supply chain. But 60-day realized vol at 120 vs BTC's 24 — the tape has crowding and leverage built in. Short-term the trade is vol; medium-term it's scarcity. (KOSPI retail forced liquidations in Samsung/Hynix add a liquidity tax.)

WDC

The TrendForce 68% number applies to NAND too — WDC is the primary US NAND pure-play. Enterprise SSD plus nearline HDD gives it a second AI storage hook. Leverage is to NAND pricing, not HBM. Cheaper way to play the memory capex absorption.

SNDK

Holder trimmed at $150 — profit-taking into strength, not a thesis break. SNDK is the pure-play NAND recovery vehicle. Key debate: are 2027 NAND price increases durable enough to justify a re-rate from here?

STX

Same profit-taking at $150 — calls trimmed into storage strength. Seagate is a capacity-storage play, not an HBM/NAND scarcity play. AI datacenter build helps, but the linkage is indirect. R/R at these levels is mediocre until the tape digests. Watch for a continuation setup.

CRBS

SemiAnalysis math is a bad omen: a 1.6T-param model costs $20M+ capex and ~1MW across 20 wafers before the first forward pass. Bull case rests on CS-6 hybrid-bonded DRAM (2029) and premium-speed decode that NVDA's high-concurrency economics can't touch. Both can be true — CRBS wins low-batch decode, NVDA wins the rest. That makes it a 2-3-year option on premium-speed token demand, not a broad AI winner.

AMD

Helios/MI455X full-stack rack co-design lands in the ballpark of Vera Rubin on total rack KW — with everything in-house (CPU, GPU, fabric, networking). MI300x matches H200 on one benchmark without the NVDA tax. But software and system ecosystem still lag; benchmark wins don't convert to share at scale. Custom ASIC design services are a call option, not a current earnings driver.

INTC

EMIB packaging won a marquee SK Hynix slot — direct beneficiary of CoWoS tightness. That's real foundry-services revenue, not roadmap promise. 18A specs are competitive on paper (Diamond Rapids: up to 256 P-cores, ~1.28GB LLC, 128 PCIe Gen6), but yields and timing are undisclosed. The foundry turnaround narrative is louder; tape-proof is still ahead.

TSM

Jalapeño proves chip design barriers are collapsing — the durable moat migrates to manufacturing. Design wins come and go; TSMC gets paid regardless. CoWoS tightness fragmenting packaging business to Intel is a competitive warning, but also a demand signal. Power bottlenecks add timing risk; structural demand intact.

QCOM

Xiaomi going in-house (Xring O3) is the Apple playbook repeating — Chinese OEMs verticalizing. QCOM needs the custom ASIC design-services pivot to offset mobile share bleed. MS model shows aggressive ASIC revenue potential, but execution and margins are unproven. Call option, not core thesis.

ARM

Every Chinese OEM in-house chip still pays Arm — Xring O3 uses Arm CPU/GPU cores on TSMC 3nm. Toll-booth beauty of the model. The offset: OEMs optimizing around Arm IP squeeze royalty per device. Challenge is monetizing AI/datacenter cores before handset-core commoditization bites.

SMTC

RECORD BACKLOG. Q3 data-center revenue guide +45% SEQUENTIAL; Signal Integrity revenue +64% y/y; ~$410M guide above expectations. Rare small-cap with hard numbers confirming the optics/connectivity upcycle. Acceleration into Q4/FY28 is the tell — tape rewarding execution, not narrative.

APLD

MS channel check: ALL construction projects on time and on budget. That directly attacks the high discount rate the market applies to AI HPC. If deliveries stay clean, project risk premium compresses. Leverage is CoreWeave demand and power access — debate is the equity discount rate, not the leases.

CORZ

Four of five CoreWeave campuses substantially complete AHEAD OF SCHEDULE — 437 billable MW. Strongest execution datapoint in the group. On-schedule delivery compresses the project risk premium. Switch/Vantage IPOs (Q4/Q1'27) provide the first real comps — likely favorable.

CIFR

Black Pearl campus delivered TWO MONTHS AHEAD of schedule. Market has been punishing all AI HPC miners with one high discount rate — delivery beats deserve spread compression. Bitcoin mining plus HPC colo hybrid adds a moving part. Keep banking the deliveries.

RIOT

AMD data center build-out on schedule AND on budget. Mining-to-HPC-colo transitions rarely stay clean. Supports the MS thesis that project execution is better than feared. Retains bitcoin mining plus AI HPC optionality. Debate: driver or capital sink?

WULF

On schedule, but total project cost now ~$9.1M/MW vs $8.6M/MW financed in October — the lone cost slipper in an otherwise clean execution tape. Still inside $8-10M guidance, but it raises equity funding needs or lowers per-MW return. Not a thesis breaker; a margin nuance.

LITE

Lumentum's strength is a forward indicator for NVDA networking revenue — optics confirming AI fabric demand, not just GPU demand. LITE is directly levered to pluggable lasers and higher-speed transceivers. The trade: long NVDA vol into earnings — optics read-through makes a networking beat more likely.

COHR

Coherent's strength alongside Lumentum says AI networking isn't waiting on NVDA's guide. COHR has broader industrial/LIDAR exposure but also high-speed optical transceiver content in AI datacenters. Key debate: does the margin mix improve as AI datacom revenue scales?

AAOI

The '10x expansion / NVDA 50% order cut' rumor is recycled fake news — debunked. US manufacturing base is real but the 10x framing is not credible. If the rumor moved the stock, it should fade. Don't chase AAOI on this noise.

GLW

Corning Glass-Bridge got caught in the same debunked rumor dump as 'CPO kills all pluggables.' Glass-based optical interconnects are a real research direction — they don't displace pluggables wholesale this cycle. No portfolio action until a named customer validates the tech.

GOOGL

TPU V7 BEATS BLACKWELL AND RUBIN ON FLOPS/WATT. Dataflow architectures are sweeping the efficiency category — Google is the quiet benchmark-setter behind the custom-silicon wave. Read-through to NVDA: custom silicon is a broad-based competitive threat, not just OpenAI. Hidden Anthropic stake is an option at the S-1 mark, not the thesis.

AMZN

Amazon is raising AI device prices on chip supply constraints and pushing customers toward rental models. Tells you AI hardware supply is tight enough to give procurement pricing power. Core read: AI capex cost inflation is real, and Amazon will price it through. Same Anthropic/OpenAI hidden-asset optionality as ZM/GOOGL at the S-1 mark.

AAPL

M5 Ultra Mac Studio supports up to 256GB unified memory (512GB in October); M6 is Apple's first 2nm chip. Apple is building the local agent inference machine — hundreds-of-billions-parameter LLMs on-device, Thunderbolt 5 clustering for pooled memory. If multiple parallel agents run at useful speeds, the consumer AI hardware cycle is underpriced. Big 'if' — but the roadmap is real.

ZM

Q2 FY27: revenue $1.28B +4.9%, adj EPS $1.55 beat, enterprise +7.8% — strongest in three years. FY guidance raised on EPS/FCF, revenue in-line; stock fell 3-4% A/H. Solid print, no AI monetization reveal — range-bound. Anthropic stake is the asymmetric variable; it becomes a hard number at the S-1.

INTU

Beat and guided FY27 growth to 9-10% vs ~14% this year — stock dropped 11%. TurboTax and Credit Karma are the best-case AI agent workflows, and AI monetization still isn't accelerating into the guide. First hard evidence AI value is concentrating in infra, not the app layer. Stock repricing to a lower terminal growth ex-AI.

BABA

HK$80BN ($10.2BN) PLACEMENT AT HK$112.70, 3X OVERSUBSCRIBED — Jack Ma bought >HK$600M, Tsai/Wu ~HK$120M near the price. Insider buying into a huge raise is a clear vote of confidence. But the quarter shows FCF outflow RMB44.7bn vs RMB18.8bn last year; capex RMB67.7bn. Even the best-funded Chinese AI player needs equity markets to fund the build.

MRNA

MRNA is trading like a meme stock on daily price action — market structure signal, not fundamental. No catalyst in the feed. Treat as noise; risk is slippage and crowded retail positioning. If you trade it, size as a momentum vehicle, not an investment.

BE

Bloom +7% pre-market after +7% intraday in a weak AI capex tape. Behind-the-meter power keeps finding bids while semis sell off. BE is the pure-play fuel cell for BTM datacenter generation — avoids grid interconnection delays. The bid is real; the question is orders vs energy scarcity positioning.

GEV

GEV (and ENR.de) are the power-equipment long theses — gas turbines, transformers, grid equipment multi-year capacity-constrained. AI datacenter TAM adds a second demand curve on top of electrification. Key debate: how much backlog is priced in after the run?

AMAT

2026-28 WFE forecast raised AGAIN on 2Q26 earnings. AMAT is the broadest semicap exposure across deposition/etch and packaging. Memory at 68% of CSP spend pulls WFE from both leading-edge logic and HBM capacity. That's a guidance raise, not sentiment — the upcycle is extending and steepening.

LRCX

WFE raised again — Lam is leveraged to NAND and advanced packaging, two of the tightest pieces of the AI memory/capex stack. The 68% CSP-capex memory number makes Lam's etch/deposition content more durable. Extends the semicap upcycle.

KLAC

WFE raised again. KLAC is the process-control bellwether, leveraged to yield ramps at leading-edge and HBM. More memory capacity and packaging complexity equals more defect-inspection dollars. High-confidence demand signal.

ASML

2026-28 WFE raised again — ASML is the litho bottleneck for leading-edge logic and memory. HBM buildout and TSMC/Rapidus competing for EUV/DUV slots keeps ASML supply-constrained. Jalapeño proving design is easier makes the manufacturing moat more valuable. Litho is the ultimate toll booth.

CDNS

Hardware emulation (Palladium-class) is the unglamorous gate between AI-designed chips and tapeout. As custom silicon design starts explode — Jalapeño and the ASIC wave — emulation capacity becomes a quiet beneficiary. AI floorplanning is real but a small slice of the workflow. Shovels to the custom-silicon gold rush.

SNPS

Same emulation/Zeus-class read-through as CDNS, plus IP and security for more surface area. AI-assisted design narrative is advancing, but the full workflow is still EDA-intensive. Quiet beneficiary of the Jalapeño design-cycle compression.

SITM

SiTime draws airtime as precision timing content per accelerator rises with rack-scale clock-domain complexity. Timing is a tiny but growing BOM line — catalyst-watching, not a confirmed datapoint. Bear case: content too small. Bull case: attach rate per socket at AI system volume.

DELL

Dell's bare-metal service anecdote — hot-spare replacement within an hour and a Dell hardware diagnosis loop — shows AI-native operators winning on service differentiation vs large cloud. Supports the case that AI server attach includes service margin, not just hardware. Also highlights competitive pressure on legacy cloud providers.

JLL

JLL data: >66GW of US datacenter capacity under construction, West Texas emerging as a leading market. Direct CRE-services beneficiary of the AI build. 66GW under construction means the supply pipeline is real and the power/interconnection bottleneck tightens.

IBM

Future Z/LinuxOne runs 2nm, 11-core with hybrid z/Arch/AArch64 plus a 2nd-gen PCIe AI accelerator — 96GB HBM3E, ~20x memory-bandwidth increase. Mainframe finally getting serious on-package AI. Targets enterprise/transactional AI — niche but defensible. Small piece of IBM's AI narrative vs Red Hat and consulting.

TRI

Thomson 1.0 scores 78.5 on legal/tax composite vs Opus's 79.5 and Gemini 3.1 Pro's 78.0 — final training run under $450K GPU. Moat is proprietary data/workflows, not architecture. Validates domain-focused AI built cheap and distributed full-weight. Fundamental proof point for data incumbents.

VLO

Amur Gas Chemical Complex blast likely delays 2.7Mt/y polymer output by 4-9 months. For VLO, the link runs second-order through Russian energy/petrochemical flows — not crude runs. Real impact: global polymer/naphtha balances. Treat as an energy-commodity input, not a VLO earnings driver.

MPC

Same Amur blast flag — weak direct linkage. The complex makes polymers, not refined products. The delay defers a small slice of global petrochemical supply. For MPC: a minor naphtha/crack distortion at most. Don't build a position on this headline.

PSX

Same Amur read-through — minimal direct petrochemical exposure. The 4-9 month polymer delay is a global supply-side datapoint, not a PSX P&L driver. If any read-through exists, it's product cracks and naphtha flows. Keep conviction low.

DINO

Same Amur blast flag — DINO is mostly refiner/midstream; the polymer delay is tangential. The event nudges global petrochemical balances, not DINO's crude-processing economics. Noise.

WMT

Walmart valuation comment is a pass-through — eventually the multiple matters to the market narrative. No fundamental datapoint in the feed. Defensive quality holding, not an AI signal. No action.

DKS

Steve Weiss recommended long DKS but the feed gives no thesis. No fundamental datapoint to analyze. Possible consumer discretionary long at a reasonable multiple — but that's inference, not source. Don't act on a bare mention.

TSLA

First-hand FSD ride drew an 'impressed' note. Qualitative product signal, not a market-moving datapoint. FSD improvement pace matters for robotaxi optionality. No financial or delivery data in source — keep conviction low.

CRDO

No new signal on the tape. Still the high-speed connectivity pure-play on AI NIC/DSP attach — wait for design-win news or a pullback.

CRWV

Quiet tape. CoreWeave remains the private comp for the AI HPC miners — lease or refinancing headlines hit CORZ/APLD/CIFR before this one.

DT

No signal. Observability gets more critical as AI agents multiply, but nothing in the feed moves the thesis.

EMR

Nothing on the tape. Emerson rides the same electrification/reshoring capex wave as GEV — no incremental datapoint here.

FROG

No fresh signal. JFrog's DevOps/software supply-chain story is quiet. No update.

KDK

No signal at all. Blank tape — can't build a view.

MDB

No signal. MongoDB remains the developer-data-platform AI beneficiary in the bull case — but nothing incremental in the feed.

MRCY

No signal. Mercury is the defense/space electronics compounder; quiet tape.

MSTR

No signal. MSTR equity is a leveraged bitcoin vol play — and BTC realized vol is compressed at 24 while MU prints 120. Nothing new in the feed.

NTNX

No signal. Nutanix private-cloud hybrid story — nothing incremental. Watch AI-infrastructure attach commentary.

OMCL

No signal. Omnicell is healthcare automation, a completely different tape. Nothing in the feed.

SBAC

No signal. Tower REITs are edge/wireless optionality at best. No datapoint.

SKHY

No direct signal, but the memory read-through is live: KOSPI retail forced liquidations in Samsung/Hynix are a liquidity tax on the whole HBM trade. SKHY is the HBM scarcity anchor — watch it as a sentiment tell for MU.

TTAN

No signal, no feed mention. Skip.

VEEV

No signal. Veeva is the vertical SaaS AI story in life sciences — quiet.

WDAY

No signal. Workday is the AI-HR agent story — closest comp to INTU's repricing if AI monetization lags. Watch for the same multiple compression.

XPEV

No signal. XPeng is EV/robotaxi, far from the AI infra tape. No datapoint.

2. Street Color / Heard (unverified)

Hearing OpenAI's talent exodus is the under-appreciated risk in the Jalapeño glow. Head of Data Centers out; CRO, COO, and Fidji Simo all gone in recent weeks. Silicon execution impresses, but retention ahead of the 2027 IPO is the new execution risk.

Word is OpenAI and Anthropic will take major equity warrants from public-company customers within 12 months — frontier tokens in exchange for warrants, starting with biotech. Novel compute-financing that moves AI capex onto customer balance sheets. First disclosure re-rates whoever holds the warrants.

Channel checks suggest Anthropic is hiring silicon engineers. Second lab going vertical on custom compute. Early read: more merchant ASIC design engagements beyond OpenAI — bull for AVGO's design-services franchise.

Hearing AVGO CDS +5bps and NVDA +1bps — both new record wides. Bondholders are done subsidizing overpriced GPUs, TPUs, and memory. Goldman's Delta-1 desk flags credit, inventory, and power as the three questions equity has ignored — that's the SOX derating from 21-22x to ~15x forward in one sentence.

Word is Meta ships an agent appstore inside WhatsApp — "WeChat mini-programs with 3B DAUs." 1P agents (personal assistant, relationship coach) plus 3P agents (food, real estate, travel, commerce). 2027 is the year of personal agents — and that optionality is not in consensus.

Word is Switch and Vantage both likely public in Q4/Q1'27. First real comps for the AI HPC miners — and the chatter says the comp will be good. That's the catalyst for spread compression across APLD/CORZ/CIFR.

Channel checks from MS's Stephen Byrd: AI HPC project execution is beating the high-discount-rate bear case. APLD all on time/on budget; CIFR's Black Pearl two months early; CORZ four of five CoreWeave campuses ahead of schedule at 437 billable MW. WULF is the lone cost slipper at ~$9.1M/MW vs $8.6M/MW financed.

Hearing the AAOI '10x expansion / NVDA 50% order cut' story is recycled fake news. Debunked, but it moved the stock. Same rumor dump tagged Corning Glass-Bridge — real research direction, not a pluggable-killer this cycle.

Word is OpenAI's Jalapeño benchmark beat compares against GB300, not Rubin. Perf/W and latency claims are legit per Richard Ho, but the basis matters. Design barriers are collapsing — the moat migrates to manufacturing (TSM/ASML).

Hearing ~60% of a market survey cites grid interconnection and permitting as a "major source of unexpected delays." ERCOT Far West congestion is the live case study — oil/gas electrification and bitcoin mining pushing against limited imports. That's why behind-the-meter names (BE, GEV) keep finding bids.

Word is the $100 ChatGPT Business Premium seat is live — SMB monetization without seat-cap friction. Steady ARPU creep, not a headline, but the installed-base monetization machine keeps turning.

Hearing Keenable (private) is the search-for-agents name to watch — ex-Yandex search team building agent-native web infra. "Crawl is like building a fab." If search-for-agents becomes an infrastructure layer, this is the scarce team.