Newsletter · · Ashutosh Agarwal

The AI Capex Tracker - June 25, 2026: Micron Beats the Bull Case, Own the Recipients

Micron beat even the bull case and the trade rotates toward the CapEx recipients over the spenders, while OpenAI unveils its Jalapeño inference chip with Broadcom, for the week of June 25, 2026.

The AI Capex Tracker

Issue: Thursday, June 25, 2026: Micron Beats the Bull Case, Own the Recipients

TL;DR

  • The binary resolved, and it resolved bullish. Micron beat even the bull case and traded +5% after the close. EPS guide of $31 vs $25 consensus, 16 strategic customer agreements (14 worth ~$100B+ each over 3–5 years), and FY27 capex guided >$40B, yet it still printed $16B of free cash flow in the quarter. The single print everyone feared went the bulls' way. (Closing Bell, Jun 24)
  • OpenAI shipped its own silicon. With Broadcom, it unveiled Jalapeño, a custom LLM-inference chip, ~50% lower cost than a typical AI GPU, designed-to-tape-out in 9 months, ramping at gigawatt scale through '27 and "full bore" in '28. Brockman: it's additive to Nvidia, "we cannot get compute fast enough." (Squawk on the Street, Jun 24; Bloomberg Tech, Jun 24)
  • The trade is the rotation: own the recipients, not the spenders. Money is piling into the CapEx recipients (memory, hardware) and away from the hyperscalers footing the bill. But the bear's arithmetic is intact, $2T of net-new revenue needed within four years to justify the build. (Closing Bell, Jun 24; The Valley Current, Jun 24)

What's new

Yesterday was the setup note, chips cracking worldwide, Micron looming "tonight." Tonight it resolved, and it resolved up. The bears who wanted a binary got one, and lost it.

1. Micron passed, and beat the bull case. Closing Bell, Jun 24, Julian Emanuel (Evercore chief equity & quant strategist) and Mehdi Hosseini (Susquehanna, semis).

Emanuel: "They beat even the bull case for almost all the metrics." The stock popped ~5% after hours, above its intraday high near $1,080. Hosseini's frame is the memory wall: "Consensus was for $25 of EPS in August and Micron guided to 31… customers have no choice but to pay a premium." The detail the market wanted: 16 strategic customer agreements, 14 of them at "at least approximately $100 billion over the remaining agreement terms" (3–5 years), that breaks the boom-bust cycle and hands Micron multi-quarter visibility. The scary line, FY27 capex "more than" $10B/quarter, i.e. $40–50B, is offset by falling capital intensity: $25B operating profit, $7B capex, $16B free cash flow in the quarter. This cycle is different because it self-funds.

"Investors want exposure to only the recipients of all that CapEx… as opposed to the companies footing the bill." (Julian Emanuel)

2. OpenAI built a chip. Broadcom built it. Squawk on the Street, Jun 24, Greg Brockman (OpenAI president/co-founder) and Hock Tan (Broadcom CEO), operators, in a CNBC exclusive.

Jalapeño is a from-scratch LLM-inference accelerator, 50% cheaper than a GPU per token, designed-to-tape-out in nine months, "blazing fast for this industry" (Brockman). Critically, it's not a Nvidia replacement: "It's very additive. We feel that we cannot get compute fast enough to keep up with demand." Nvidia stays OpenAI's training partner; Jalapeño targets inference. Tan's framing is the structural call: "every… frontier model developer will eventually go towards building their own silicon," and demand from his six hyperscaler customers is "simply insatiable… not just '26, not '27. We're seeing that same and even elevated demand in '28." Deployment: small prototype in '26, scaling '27, "full tilt first half '28." (Bloomberg Tech, Jun 24: Microsoft added $41B of data-center commitments in Q2 ($197B total); hyperscalers have committed $850B in future leases.)

3. CJ Muse: memory tightens more in '27, and TSMC picks the winners. Squawk on the Street, Jun 24, CJ Muse (Cantor Fitzgerald, semis). "Supply is going to be even tighter in '27 than '26," so earnings power extends into '28. Bulls model $200 of Micron EPS next year (a 5x multiple). His allocation tell: "all of our work suggests TSMC is prioritizing NVIDIA, AMD, and Broadcom… those are your three winners on the custom-silicon front." Marvell's AWS + Microsoft + optics is driving "$10 [EPS] into 2028."

4. The power read is GE Vernova: sold out to 2029. Squawk on the Street, Jun 24, Seema Mody live from GEV's South Carolina turbine plant (operator color). GE Vernova is sold out of gas turbines through 2029, with orders booked to 2031; turbine pricing is up 300% in three years. Microsoft just bought seven to power a 2.7 GW Texas project, and "executives from every hyperscaler, including OpenAI['s] head of power, have walked the floor." Faber adds two more hyperscaler gas deals are coming before Labor Day, similar to the 2.67 GW Microsoft–Chevron deal. Behind-the-meter gas is winning because "they need a standalone power option. They can't just rely on the grid."

5. Cerebras' first public print: -15% on the margin guide. Squawk on the Street, Jun 24, Andrew Feldman (Cerebras CEO). Record $191M revenue and a raised full-year margin guide, yet shares fell ~15% on lockup mechanics and a near-term dip from "renting back" gear to feed OpenAI demand. The structural tell: Cerebras "sidesteps" all three supply chokepoints, HBM, TSMC's CoWoS packaging, and the 3nm node, running SRAM on 5nm.

The debate

Bull steel-man, the recipients are printing, and demand is real. Micron beat the bull case with 14 customers committing $100B+ apiece; supply tightens further in '27 (Muse); demand is "insatiable" through '28 (Tan) and OpenAI "cannot get compute fast enough" (Brockman). And it self-funds: $16B FCF in the quarter even against a $40B+ capex guide. The rotation into memory/hardware is rational, that's where the earnings power actually shows up. (Closing Bell, Jun 24; Squawk on the Street, Jun 24)

Bear steel-man, the demand side is two insolvent customers and a circular loop. The cleanest bear math on the tape: AI infrastructure needs $2T of net-new annual revenue within four years to pencil, while 89% of AI-startup revenue flows to just two unprofitable entities (OpenAI, Anthropic) and 61% of Nvidia's revenue comes from four customers, "if just one… decides they've purchased enough… the ripple effect would be violent." Money funds Nvidia → startups buy Nvidia → clouds fund the labs → labs rent the clouds back. (The Valley Current, Jun 24) Even the bulls hedge: Hosseini flags the 2028 training-to-inference transition "could be challenging," and a FY27 model of $142B net income on $207B revenue (69% net margin) invites Emanuel's question, "can we bank on that being… sustainable?" (Closing Bell, Jun 24)

Sell signals to watch: any hyperscaler using "matching demand" or trimming 2-year capex at July earnings; a second producer confirming the SK Hynix HBM throttle from yesterday; token-price-driven capex moderation (the rotation is already sniffing it); Oracle CDS spikes or accelerating private-credit redemptions; a major enterprise (Uber) publicly capping AI spend.

Stocks in play

NVDA. Bull: still OpenAI's training partner, "additive" to Jalapeño, and a confirmed TSMC-priority winner (Muse). Bear: back below $201 on the day, now visibly encircled on inference by Jalapeño, Google TPU, and Amazon Trainium. Next: Rubin ramp; Q2 print. (Squawk on the Street, Jun 24)

AVGO. Bull: Jalapeño is the headline, Tan's six customers are "insatiable" through '28 and "every frontier lab" will build custom; a TSMC-priority winner. Bear: OpenAI volume is back-end loaded, "small prototype" in '26, full tilt only 1H'28, and Tan conceded Microsoft must backstop scale. Next: custom-silicon ramp milestones; '28 deployment proof. (Squawk on the Street, Jun 24; Bloomberg Tech, Jun 24)

AMD. Quiet on the tape, no MI450X/Helios signal, though Muse names it one of three TSMC-priority custom-silicon winners. Next: AMD Advancing AI Day, July 2026. (Squawk on the Street, Jun 24)

MSFT. Bull: added $41B of data-center commitments in Q2 (~$197B total) and is the named backstop for Jalapeño-scale deployment. Bear: the archetypal "spender" the market is rotating away from. Next: FY26 Q4 capex commentary at July earnings. (Bloomberg Tech, Jun 24)

GOOGL. Bull: TPU is now the validated template, Muse calls Jalapeño "very similar to what Google is doing with their TPU." Bear: still a "spender" footing the bill, lagging the recipient rotation. Next: July capex guide. (Squawk on the Street, Jun 24)

AMZN. Bull: Trainium template validated by the Jalapeño move; Marvell's AWS business is a core driver of Muse's "$10 EPS into 2028." Bear: a "spender," and flagged in recent issues as next to tap debt markets. Next: July earnings. (Squawk on the Street, Jun 24)

META. Quiet on the tape, cited only as a "spender" that's fallen from ~$800 toward ~$565. Treat as a carry position into July earnings. (Closing Bell, Jun 24)

Read-throughs

  • Memory / HBM (MU), the recipient trade got its green light. Micron beat the bull case, has HBM sold out through end-2026 at ~81% gross margin and a fresh Anthropic deal, and now carries $100B-class multi-year contracts. SK Hynix is reportedly weighing a ~$29B U.S. IPO (~$1T cap), a potential re-rate read for the whole complex. Own the recipients. (Motley Fool Hidden Gems, Jun 24; Squawk on the Street, Jun 24)
  • Power / thermal (VRT, ETN, GEV), GE Vernova is the cleanest actionable signal on the tape. Sold out to 2029, orders to 2031, pricing +300% in three years, Microsoft buying seven turbines for 2.7 GW, and two more hyperscaler gas deals before Labor Day. Vertiv and Eaton were quiet, but underwrite them off this demand curve, behind-the-meter gas is the build of record. (Squawk on the Street, Jun 24)
  • Optics / networking, Marvell is the name. Muse's "$10 EPS into 2028" rests on AWS + Microsoft silicon plus "tremendous strength in optics." Astera/Credo/Coherent/Lumentum quiet this cycle. (Squawk on the Street, Jun 24)
  • Custom silicon, TSMC allocation is the gating asset. Jalapeño + TPU + Trainium means every frontier lab is now building. The constraint isn't ideas, it's wafers, and they flow to NVDA, AMD, AVGO. That trio is the structural long; the merchant-GPU-only thesis narrows.

What changed vs last issue

Last issue (Jun 24, "Chips crack worldwide. Micron's the verdict tonight.") was a setup note into a global memory unwind. In 24 hours, the setup became the verdict, and it broke bullish:

  • Micron passed. It beat the bull case (+5% AH), with 16 customer agreements and $40B+ FY27 capex funded by $16B of quarterly FCF. The "could fall 30% and stay in an uptrend" air-pocket setup didn't fire; the correlated unwind didn't follow through.
  • The custom-silicon war escalated, OpenAI now has a chip. Jalapeño (with Broadcom) is the biggest structural development since the Blackstone-TPU deal: every frontier lab building, additive to Nvidia, gating on TSMC wafers.
  • Power got a hard print: GE Vernova sold out to 2029, Microsoft's seven turbines, two more gas deals by Labor Day.
  • The framing flipped to recipients > spenders (Emanuel), the cleanest single positioning takeaway of the week.
  • Bear math sharpened from yesterday's depreciation/Stargate thread to Valley Current's $2T revenue gap and 61%-of-Nvidia-from-4-customers concentration. Yesterday's SK Hynix HBM-throttle remains single-sourced and unconfirmed. No new ERCOT/SB6 print this cycle.