Newsletter · · Ashutosh Agarwal

Meta Warns on AI Demand as Challenger Silicon and Memory Costs Take Center Stage - AI Accelerators: GPUs, Custom Silicon & Optics - Week of June 29, 2026

AI accelerators newsletter for the week of June 29 to July 6, 2026. Meta's plan to sell excess compute became the bears' exhibit, challenger silicon (Groq, Etched, Qualcomm) supplied the only fresh operator detail while Nvidia and AMD stayed dark, and the memory shortage bled downstream into device price tags.

AI Accelerators: GPUs, Custom Silicon & Optics

Week of June 29, 2026: Meta Warns on AI Demand as Challenger Silicon and Memory Costs Take Center Stage

For five issues the accelerator tape argued about supply. This week it flipped to demand and financing: Meta became the exhibit, challenger silicon (not Nvidia, not AMD) supplied the only fresh operator detail, and the memory shortage moved from a chart to device price tags.

1. From "How Much Can They Build?" to "Who's Buying, and Who's Funding It?"

The pivot was Meta's reported plan to sell excess AI compute. The bear read from a guest on Prof G Markets, "The AI Trade Just Got A Warning From Meta" (Jul 2) (pundit): "80% of all AI compute is owned by or used by OpenAI or Anthropic, with the rest… being Meta." Excess to lease means "a mirage of demand… leaving scraps for the rest of us," and the tell is "you know that we're at the end when Anthropic buys compute from Meta." He pegs Oracle's Stargate Abilene (~880 MW critical IT) at ~$10B/yr, calls the "$20 billion a gigawatt" projections "ludicrous," and says of "over 100 gigawatts… under construction," he finds demand for ~6, one bear's estimates, not audited, but the direction is the story.

The plumbing gave bears more. On Eurodollar University, "Private Credit Just Burst The $25 Trillion AI Bubble" (Jul 5) (pundit, citing Bloomberg): Blue Owl's credit fund saw ~$4.2B of redemption requests in Q1 and ~$3.6B in Q2, its tech-income fund was hit for 38.1% of shares in a quarter, ~$14B is "trapped" across private-credit vehicles, and Blackstone is selling data-center stakes and abandoning a Virginia project via QTS. The regime change that matters: "the marginal lender says… show me the signed lease, where you're getting the power, and show me your exit, the entire growth curve changes… much of it is really a debt story."

The circularity got more explicit. On RiskReversal, "David Rosenberg" (Jul 1), host Dan Nathan detailed a ~$35B SPV set up in early June by Blackstone and Apollo to finance Google TPU purchases (Broadcom-designed) for Anthropic, atop Nvidia "funding tens of billions" to its own customers, a backstop that resurfaced as Nvidia's "AI Compute Partnership" program on The Information's TITV (Jul 2). Rosenberg (pundit) likened it to "the late 1990s… when everybody was each other's customer," and says that adjusting for the depreciation skew, the S&P's real P/E "is actually 30." OpenAI, separately, reportedly paused its IPO after failing to clear a $1T mark (Prof G Markets, Jun 30).

The bull rebuttal was, unusually, quantitative. Summarizing an Exponential View study, The AI Daily Brief, "How Big Is the AI Economy?" (Jun 30) put hyperscaler/neocloud capex at $848B in 2026 and made the payback case: quarterly revenues have exceeded capex depreciation since Q4 2025, and older GPUs earn yields "into year 7, 8, and even 9" despite six-year depreciation, the direct counter to the cliff thesis. The catalyst that settles it: Avory (Jul 2) flags a "critical 6-8 week window": FY27 capex guides (hyperscalers spend ~90% of operating cash flow; ~$785B in 2026, nearing $1T in 2027) land on late-July/early-August calls. Any flattening is the tell.

2. Challenger Silicon Speaks, While Nvidia and AMD Stay Dark

With Nvidia's Blackwell/Rubin roadmap and AMD's MI-series both silent (see negative space), the only fresh operator chip detail came from challengers. Groq founder Jonathan Ross, who "created Google TPU," walked through his ~$20B Nvidia partnership (pairing Groq's LPU with GPUs, since "GPUs and LPUs combined ended up giving better performance across the performance curves") on David Senra (Jul 5) (operator). Origin: "we had gone to Jensen asking, could we buy about 100,000 GPUs… and Jensen saw what we had done and thought maybe it would be better to make this available to all of their customers." Candidly: "As the creator of the TPU, I have to admit, GPUs are now better."

On Invest Like the Best, "Etched" (Jun 30) (operators), the founders put numbers on inference-only silicon: model-flops utilization on GPUs runs "between 20% and 50%"; their chip runs "under half the voltage of any other AI chip"; and where a Blackwell chip-to-chip hop takes "about 4,000 nanoseconds," their custom interconnect cuts latency "by more than a factor of 5x." Tying it together, Training Data with Dylan Patel (Jun 30) (operator/analyst), co-design is "AI's real 100x." New detail: Google runs three distinct TPU architectures, "a TPU with Broadcom… a different architecture than the TPU with MediaTek," plus a third; "in the case of Google, hundreds of billions of dollars a year of their own ASICs." Qualcomm's data-center debut also kept surfacing (The Six Five, Jun 29; AI Inside, Jul 1): custom silicon for Microsoft and Meta, $5B FY27/$15B FY29 targets, and a Meta CPU deal from 2028 (reported, not company-confirmed on-air).

3. Memory: The Shortage Bled Downstream

Issue 005 covered take-or-pay contract mechanics; this week the story moved into everyone else's cost structure. On Daybreak (Jun 30), Nothing CEO Carl Pei (operator) said "memory is now the most expensive component in a smartphone… more than 50% of the total hardware bill," with data centers set to eat >70% of high-end memory output in 2026; Nothing cancelled budget models. Apple hiked Mac/iPad prices up to $200+, Tim Cook saying he'd "never seen a component price increase this much this quickly." Analysts piled on via Tech Brew Ride Home (Jun 29): Jefferies models memory +40-50% in Q3 and +30-40% in Q4, and GoPro's memory costs are up 80-115%; The Circuit (Jun 29) pegged HBM +60% QoQ, >300% annually. One non-obvious overlay from Real Vision's Macro Mondays (Jun 29): Andreas Steno (pundit) notes DRAM spot keeps climbing with no drawdown even as the token-expenditure index rolls over, partly demand front-loading on helium-supply worries. The Micron print (~$42B revenue, ~85% gross margin, relayed across several shows) is backdrop; the fresher signal is that the shortage is now a tax on every device maker.

4. Power: The Binding Constraint Is Now the Interconnection Queue

Where Issue 005 framed power as Bloom-vs-Microsoft, this week's operator reframed it as a queuing problem. On Open Circuit (Jul 2), Digital Realty's Ian Black (operator) said the sector "feels like the world of renewables in 2014," utilities are "cutting 80 percent of applicants" with financial-security rules mirroring MISO/PJM circa 2013-14, and "powered land… is a much higher bar than it was six or nine months ago. You may have to have permits in hand before you can get studied." He's "waiting on Oncor for two years… in Texas," yet signed a separate ESA "in 20 days" when it was "entirely predicated on BYO power." Columbia Energy Exchange (Jun 30) added the mechanism: Texas now charges $55,000/MW to enter the queue, and data centers are being pushed into "controllable load" status, PJM is already under emergency DOE curtailment orders.

Negative Space: What Didn't Get Said Is Signal

  • Optics runs on pundit enthusiasm, not operators. Limitless (Jul 1) cited Lumentum units rising ~20M to ~60M YoY and Jensen's ~$2B stake; MoneyFlows (Jul 2) relayed Jensen's line that "the silicon photonics… capacity we need is substantially higher than the world has today." But no Coherent or Lumentum executive was on the tape, a mirror of Issue 004's Coherent-CEO lead.
  • AMD is absent for a third straight issue, a dedicated MI350/MI400/Lisa Su sweep returned zero. Conspicuous for a supposed #2.
  • Nvidia's roadmap stayed quiet. Bryan Catanzaro was on The MAD Podcast (Jul 2), but on open models (Nemotron), not the Blackwell/Rubin ramp.

Bottom line: bulls and bears are now arguing about the same six-to-eight weeks. If the late-July/early-August FY27 capex guides hold, the payback math wins the round. If they flatten (with private credit tightening and Meta dumping "excess" compute) the tape above becomes the story of a top, not a boom.