Newsletter · · Ashutosh Agarwal

Micron's Blowout Reframes Memory as a Scarcity Asset as the AI Bubble Debate Sharpens - Weekly Semis & AI Infrastructure Podcast Recap - Week of June 28, 2026

Semiconductors and AI infrastructure podcast recap for the week of Jun 22-28, 2026. Micron's fiscal Q3 blowout reframed memory as a scarcity asset, OpenAI and Broadcom unveiled the Jalapeño custom inference chip, and the AI-bubble chorus grew louder over roughly $700B of hyperscaler capex.

Weekly Semis & AI Infrastructure Podcast Recap

Week of June 28, 2026: Micron's Blowout Reframes Memory as a Scarcity Asset as the AI Bubble Debate Sharpens


Window: Jun 22-28, 2026

Top of mind this week

Micron's fiscal Q3 blowout on June 24 was the gravitational center of the week: revenue near $41.5B (roughly 4x year-over-year), gross margins of ~85% (versus ~39% a year ago and a historical peak near 60%), and 14 to 16 long-term strategic agreements locking in ~$100B of minimum guaranteed revenue each through 2029 to 2030. The story radiated outward: memory inflation forced Apple, Microsoft and automakers to raise prices, while the central debate hardened into "structural re-rating versus cyclical peak." Running alongside it were two other big arcs: OpenAI and Broadcom's "Jalapeño" custom inference chip (a hedge away from NVIDIA) and a louder AI-bubble chorus warning that ~$700B of hyperscaler capex is producing an "earnings bubble."

Dominant themes

  1. The memory super-cycle and the secular-vs-cyclical fight. Micron printed records: EPS $25.11 (+1,215% YoY), Q4 guided to $49-51B revenue and $30-32 EPS, DRAM ASPs +60% and NAND +80%, data-center revenue jumping from ~$1.5B annually to ~$11.5B in a single quarter, driven by price not volume. Bulls (Evercore's Julian Emanuel, Susquehanna's Mehdi Hosseini) argue take-or-pay contracts through 2029 turn memory into an "irreplaceable resource" (Closing Bell); bears (Morningstar's Will Kerwin) note the stock trades ~2x a $455 fair value and that six memory makers will add supply into late 2027 to 2028 (The Morning Filter).

  2. Memory inflation hits consumers and OEMs. Tim Cook confirmed memory costs are the primary driver of Apple price hikes; reported increases span MacBook/iPad ($100-500), iPhone 18 Pro (potentially $1,299), Mac Studio M3 Ultra (+$1,300), Vision Pro (+$200), Xbox ($100-150). 128GB DDR5 reportedly went from ~$800 to ~$2,900 in a year (Motley Fool Hidden Gems); GM, Ford and Honda each flagged hundreds of millions in losses from the DRAM crunch (Automotive News Daily Drive). Apple is reportedly exploring Chinese suppliers YMTC and CXMT at up to 30% discounts.

  3. Custom silicon accelerates (OpenAI-Broadcom "Jalapeño"). A from-scratch LLM-inference ASIC designed in ~9 months with Broadcom (and MediaTek), targeting gigawatt-scale deployment later this year, with claims of ~50% lower cost / better performance-per-watt versus current GPUs and 8 HBM stacks. Broadcom's Hock Tan: "every frontier lab will develop custom chips" (Squawk on the Street). Read as an inference-cost hedge against NVIDIA rather than a frontal assault.

  4. Hyperscaler capex and the power bottleneck. 2026 spend estimates clustered around ~$700B (Amazon ~$200B, Alphabet ~$175-185B, Meta ~$115-135B, Microsoft ~$110-120B), +60% YoY, the largest concentrated capital sprint since the railroads. Power is "the gating issue": Microsoft's 20-year Chevron natural-gas deal anchors the 2.67GW Project Kilby in West Texas (GE Vernova turbines), with more hyperscaler power deals expected before Labor Day (Squawk on the Street). Financing is increasingly off-balance-sheet (SPVs, private credit; Meta ceded ~80% equity in Hyperion to Blue Owl).

  5. Networking, optics and the "picks-and-shovels" trade. Optical interconnect is framed as the next bottleneck as clusters outscale copper (silicon photonics ~5-8 pJ/bit vs 15-25 for electrical); Coherent (NVIDIA-backed, $2B) and Lumentum are levered to a 2028 buildout (Chip Stock Investor). Coatue's Philippe Laffont and GMO favor equipment suppliers (ASML, AMAT, LRCX, KLA) that win regardless of which chip design prevails (Excess Returns).

  6. The AI-bubble debate got louder. BCA's Peter Berezin calls it an "earnings bubble" (hyperscalers heading toward $2.5T of AI assets and ~$500B annual depreciation by 2030) (NAB Morning Call); Lance Roberts warns "the blow-up is going to be spectacular," noting ~$760B of capex announced vs ~$211B actually expensed (Thoughtful Money); GMO likens $700B (2.2% of US GDP) to late-1990s fiber. Counterpoint: today's leaders (NVIDIA, Alphabet) have real cash flows, unlike the dot-com era.

Key debates

  • Memory: structural re-rating vs cyclical peak. Emanuel and Hosseini (beat even the bull case; multi-year visibility from $100B contracts) and Pomp guest Jordi Visser (Micron to $2T market cap within a year) versus Morningstar's Will Kerwin (2x fair value, supply response coming 2027 to 2028) and The Real Investment Show (memory lacks NVIDIA's GPU monopoly moat; 80%+ margins attract competition, per The Real Investment Show). Bloomberg Businessweek framed the exact question: TSMC-like re-rating or cyclical peak?
  • Jalapeño: NVIDIA threat or just inference cost optimization? Bloomberg Tech / Squawk lean bullish on the custom-silicon shift; TechCrunch's Equity and 20VC lean toward "hedge, not replace," with 20VC questioning whether vertical integration distracts OpenAI ("made in a different era of abundance").
  • Hyperscaler capex: durable demand vs negative FCF / cheaper models. Wall Street Unplugged and FinPod see a durable $700B+ cycle; Fast Money flagged that Microsoft could cut from $185B toward $120B if cheaper models (DeepSeek V4 at ~$0.87/M tokens vs Anthropic ~$50/M) compress economics; TFTC argues hyperscalers are falling behind frontier labs despite the spend.
  • NVIDIA: cheap or part of the bubble? Laffont calls NVIDIA undervalued at 13-14x forward (Squawk Pod); Berezin and technician Gareth Soloway see it as the centerpiece of the bubble / in a technical downtrend.
  • China supply chain. Grace Shao (Odd Lots) on Chinese labs re-engineering DeepSeek V4 onto Huawei chips for independence (Odd Lots); alleged ASML EUV smuggling (denied); Apple courting YMTC/CXMT, all pressure the incumbent memory and equipment narrative.

Stocks discussed with bull/bear angle

Ticker Direction Source / Speaker Argument
MU Bull Closing Bell (Emanuel, Evercore; Hosseini, Susquehanna) Beat even the bull case; "memory wall" + 16 agreements ~$100B each give multi-year revenue visibility
MU Bull Bloomberg Businessweek Q4 guide $50B vs $43.2B; 86% GM; HBM4 shipping in volume; lower-power HBM edge
MU Bear The Morning Filter (Will Kerwin, Morningstar) Trades ~2x $455 fair value; six makers add supply late 2027 to 2028, cyclical pressure returns
MU Bear/cautious The Real Investment Show Memory is cyclical and lacks NVIDIA's monopoly moat; 80%+ margins invite competition
AAPL Bear Motley Fool Hidden Gems Tim Cook: memory costs primary culprit in price hikes; DDR5 128GB ~$800 to ~$2,900 YoY
AVGO Bull Squawk on the Street (Hock Tan; Greg Brockman) Jalapeño design win; "every frontier lab will develop custom chips"; gigawatt-scale deployment
NVDA Bull Squawk Pod (Philippe Laffont, Coatue) Trades 13-14x forward; undervalued despite custom-silicon competition
NVDA Bear NAB Morning Call (Peter Berezin, BCA) "Earnings bubble"; when depreciation hits hyperscalers, GPU demand slows
NVDA Bear (China) Odd Lots (Grace Shao) China deliberately reducing NVIDIA dependence; DeepSeek V4 re-engineered onto Huawei chips
QCOM Bull Schwab Network (QCOM CFO & COO) $15B 2029 data-center target (from $5B); Meta and Microsoft as anchor custom-chip customers
MSFT Bull Squawk on the Street 2.67GW Project Kilby secured via 20-yr Chevron gas deal; behind-the-meter cost advantage
MSFT Bear/caution Excess Returns (GMO) $200B DC spend; unsustainable earnings bubble as depreciation lags capex
GEV Bull Squawk on the Street Turbine supplier for Microsoft's 2.67GW Project Kilby
CVX Bull Bloomberg Intelligence Monetizing stranded Permian gas via 20-year Microsoft power contract
COHR Bull Chip Stock Investor NVIDIA-backed ($2B); InP transceivers as optical networking bottleneck eases
LITE Bull The Deep Edge Podcast Silicon photonics 3-5x more power-efficient than copper; optical is the scaling bottleneck
AMAT / LRCX / KLAC / ASML Bull Excess Returns (Ian Smith) "Picks and shovels" win regardless of which AI chip design prevails
TSM Bull BBC Business Daily 70% foundry share, 90%+ of advanced AI chips; structural chokepoint
GM / F / HMC Bear Automotive News Daily Drive (John Bozzella) DRAM +450% in 4 months; each reporting hundreds of millions in losses
DELL / HPE Bear Business of Tech Traditional x86 server market constrained by memory shortage; margin pressure
Cerebras Bear The Rundown (Zaid Admani) Fell 10%+ despite beat; operating margins -28 to -32%; 37% GM vs NVIDIA 70%+
Semis (broad) Bear Thoughtful Money (Lance Roberts) Parabolic cyclical bubble; $760B capex announced vs $211B expensed; "blow-up will be spectacular"