Disk-Delay Tariff: The Hidden Surcharge on Every AI Query
Back the flash kings and rack-smiths; short the vendors still monetizing milliseconds.
1. The Temperature Shift
AI has converted storage from a bulk-capacity problem to a latency-and-throughput race. New language-model clusters stream terabytes per second; if data trickles in at millisecond HDD speeds, expensive GPUs sit idle. Flash and high-bandwidth memory respond in microseconds and, in dense packages, slash rack-count and power bills by as much as 80%. 232-layer QLC SSDs have already cut the raw $ / TB premium over disks to roughly 3×, and industry analysts now peg outright cost parity, before power and cooling, around 2026. Meanwhile, HDD makers’ capacity road-maps (40 TB HAMR) do nothing to narrow the latency gap and reach scale only after 2026. Cloud buyers have noticed: after last year’s GPU binge, they are funneling fresh cap-ex into all-flash tiers, NVMe fabrics, and 800G Ethernet that keep accelerators busy.
2. Long Book: Five Ways to Monetize the Heat
Pure Storage (PSTG). A pure-play on all-flash arrays. Q3 FY25 revenue rose 9% YoY as the firm disclosed a top-four hyperscaler licensing its 75 TB DirectFlash Modules, with 150 TB parts sampling later this year, management still says “no new HDDs after 2028.”
Micron (MU). Supplies the bits, HBM3E for GPUs and 232-layer QLC NAND for hyperscale SSDs. Fiscal Q3 2025 revenue hit a record $9.3 B; data-center sales now exceed half of total revenue, and HBM grew ~50% sequentially.
Super Micro (SMCI). Racks the whole AI stack, GPU trays, petabyte NVMe shelves, liquid cooling. Q2 FY25 revenue jumped 54% YoY and management said “AI platforms accounted for over 70% of the quarter.” They guide $23.5-25 B for FY25 despite tariff noise.
Arista Networks (ANET). Builds the Ethernet fabric that feeds flash and GPUs. Q1 2025 revenue climbed 27.6% YoY; CEO Jayshree Ullal reiterated a $1.5 B AI-networking run-rate for 2025 alongside a target of $8.2B total revenue.
NetApp (NTAP). The flash conversion story. FY-2025 all-flash array run-rate hit $4.1B (+14% YoY) while public-cloud services jumped 43%. Two-thirds of hybrid-cloud revenue already comes from all-flash ONTAP.
3. Short Book: Five Assets Stuck in the Cold
Western Digital (WDC). As of Feb 2025 the flash unit is formally spun out; what remains is an HDD pure-play just as hyperscalers pilot QLC for deep-archive tiers.
Seagate (STX). Has shipped 40 TB HAMR samples but says volume quals extend into 2026; density rises, latency doesn’t. Revenue still 44% below FY 22 despite a recent bounce.
Intel (INTC). Wrote off $559m to exit Optane in 2022 and sold its NAND fab; as cloud budgets swing toward GPUs, Xeon sockets lose wallet share.
Silicon Motion (SIMO). A commodity SSD-controller vendor. Q1 2025 sales fell 13% QoQ / 12% YoY as major NAND makers moved to in-house PCIe Gen5 ASICs.
Himax (HIMX). Smartphone/TV display-driver house with no AI infra angle; large-panel driver revenue shrank 28% YoY in 2024 and guidance calls for flat-to-down 2025.
4. Model Portfolio Allocation: 2% Gross Weight
We run the basket at 2% gross, 1% long, -1% short.
Long sleeve (1%)
• Pure Storage (PSTG) 0.25%
• Micron (MU) 0.25%
• Super Micro (SMCI) 0.20%
• Arista Networks (ANET) 0.15%
• NetApp (NTAP) 0.15%Short sleeve (-1%)
• Western Digital (WDC) -0.25%
• Seagate (STX) -0.25%
• Intel (INTC) -0.20%
• Silicon Motion (SIMO) -0.15%
• Himax (HIMX) -0.15%
Weights lean on the pure-plays (PSTG/MU vs. WDC/STX) and trim exposure where idiosyncratic risks are higher or beta is lower (NTAP, HIMX).
5. What Could Go Wrong?
HDD miracle. A sub-$10/TB 60 TB disk that ships at scale could slow flash cannibalization.
NAND glut. Oversupply would pressure PSTG/MU margins, though it would also hammer WDC’s remaining flash JV.
Cap-ex pause. A macro shock might stall AI build-outs. The paired shorts temper absolute downside.
Execution blips. SMCI supply chain hiccups or ANET losing deals to InfiniBand could cap long upside.
Conclusion
Long the media that keeps GPUs fed; short the platters, controllers, and CPUs that can’t keep up. Until hard disks beat physics, or AI loses its appetite, this thermocline shift should keep rewarding the hot side and melting the cold.
Disclaimer:
Ridire Research is an independent research publication operated by Ridire Research LLC and affiliated with Ridire Capital Management LLC (“Ridire Capital”), an Exempt Reporting Adviser under the U.S. Investment Advisers Act of 1940. Ridire Research is not registered as an investment adviser and does not provide personalized investment, legal, accounting, or tax advice.
Informational & Educational Purpose Only
All materials—including text, charts, model portfolios, and explicit labels such as “Buy,” “Sell,” “Hold,” “Long,” or “Short”—are published solely for general informational and educational purposes. They reflect the author’s views at the time of writing, derived from publicly available data, proprietary frameworks, and market analysis, and are not tailored to any reader’s specific objectives, financial situation, or risk tolerance. Subscription to this Substack does not create an adviser–client relationship with Ridire Capital or its principals.
No Offer or Solicitation
Nothing herein constitutes (i) an offer to sell or the solicitation of an offer to buy any security or other instrument, (ii) a recommendation to participate in any investment strategy, or (iii) a solicitation for investment advisory services. Any references to trades, allocations, or vehicles should be viewed as hypothetical model illustrations only. Offers, if ever made, will be made solely by confidential offering documents and only to qualified investors in jurisdictions where permitted.
Potential Conflicts of Interest
Ridire Capital, its affiliates, employees, and related accounts may hold, increase, decrease, initiate, or exit positions, long or short, in securities or digital assets discussed, without notice or obligation to update disclosures. Such positions may be inconsistent with views expressed in this publication. Ridire Capital, related entities, and personnel may initiate, modify, or exit positions in any mentioned security before or after publication without further notice. We maintain internal policies, including trading blackout windows and conflict reviews, to mitigate potential conflicts of interest.
Accuracy & Forward-Looking Statements
Although we strive for accuracy and analytical rigor, information may become outdated and may contain errors or omissions. Forward-looking statements, projections, or target prices are inherently uncertain and may differ materially from actual results. No warranty, express or implied, is given as to completeness, accuracy, or reliability.
Risk Acknowledgment
Investing involves substantial risk, including the potential for complete loss of capital. Past performance, whether actual, indicated by back-tests, or modeled, is not indicative of future results. Securities, derivatives, and digital assets mentioned may be illiquid, highly volatile, or subject to regulatory change.
Reader Responsibility
Readers should conduct their own due diligence, consider their personal circumstances, and consult a licensed financial professional before acting on any information contained herein. By reading this publication you agree that Ridire Research LLC, Ridire Capital Management LLC, and their affiliates accept no liability for any direct or consequential loss arising from reliance on the information presented. This research is not directed at persons in jurisdictions where such distribution would be contrary to local law.
The bear case on WDC makes sense for cold storage tiers but I wonder if you're underweighting the nearline opportunity. A lot of hyperscalers are still using HDDs for warm data where latency isnt as critical and the cost differntial still matters. Also the post spinoff WDC might actualy be better positioned to focus on that niche rather than fighting the NAND commoditization battle. Not saying the short is wrong but curious how you weigh those factors.