The AI Stack Weekly

Issue 11 · Week 27 of 2026.

July 4, 2026/Industry brief · ~7 min read/Public sources onlyDownload brief

The Bottom Line

The sell-side of AI compute gained two hyperscale balance sheets in 48 hours — just as the model layer started cutting prices.

Flywheel arcAll three lenses

Within two days, Bloomberg reported Meta is building "Meta Compute" to sell excess AI capacity (and possibly hosted Muse Spark access), knocking 12-15% off CoreWeave and Nebius in a morning, and SoftBank answered with SB Neo, a Delaware neocloud targeting 10GW by roughly 2030. The compute-rental market just gained two entrants with balance sheets larger than the entire listed neocloud category — exactly as the Epoch analysis showing hyperscaler free cash flow hitting zero around Q3 2026 became the consensus macro frame. The software lens ran the same economics in the other direction: Anthropic launched Claude Sonnet 5 at $2/$10 introductory pricing, and OpenAI's gated GPT-5.6 preview promises a Terra tier at half GPT-5.5's cost — the model layer is cutting prices into the very demand that is supposed to fill all this capacity. Meanwhile the frontier itself became a regulatory variable: GPT-5.6 previewed to only ~20 government-vetted organizations, while Claude Fable 5 returned globally on July 1 after export controls were withdrawn. Net/net: the industry's central question moved from "who has the best model" to "who can profitably fill capacity they have already financed" — boards should stress-test compute-vendor concentration, investors should separate scarce-input owners (memory, power) from capacity resellers, and architects should use the price war to renegotiate inference contracts now.

The three lenses

What moved this week, and what to do about it.

12 events across the flywheel — 5 software, 4 hardware, 3 networking.

Software.

Jun 30
Anthropic launched Claude Sonnet 5 — its most agentic Sonnet, the new Free/Pro default — at $2/$10 per MTok introductory pricing through Aug 31 (then $3/$15)
Anthropic newsroom; TechCrunch
Jun 30-Jul 1
Claude Fable 5 was restored globally after export controls were withdrawn; Mythos 5 access returned for ~100 US critical-infrastructure organizations on the Commerce 'Annex A' list
Anthropic newsroom
Jun 26
OpenAI previewed the GPT-5.6 family (Sol $5/$30, Terra $2.50/$15 at ~GPT-5.5 performance, Luna $1/$6) — limited to ~20 vetted partner organizations at the US government's request, broad GA 'in the coming weeks'
OpenAI; Reuters syndication; The Batch
Jul 2
GitHub shipped Copilot agent session streaming to public preview — enterprise admins can stream all agent session data (cloud agents, CLI, IDEs) to SIEM endpoints including Microsoft Purview
GitHub Changelog
Jun 30-Jul 2
Gemini 3.5 Pro missed its implied June GA window; no public API model ID as of Jul 2, with reporting pointing to a July target while 3.5 Flash carries the load
TokenMix status check; Business Insider via Pondero

What this means

The software story is a price-and-permission story: closed labs are cutting rates into open-weight cost pressure while Washington now sits inside the release pipeline itself. Architects should use Sonnet 5's intro pricing and the promised Terra tier to renegotiate inference contracts this quarter — and treat model availability as a regulatory risk to engineer around, with the Model Pulse carrying the full architecture read.

Hardware.

Jun 30
SK hynix filed for a Nasdaq ADS listing seeking ~$29.4B (ticker SKHY, ~Jul 10 debut expected) — potentially the largest ADR offering in history, with proceeds earmarked for fab and EUV capex
SEC filing via IBTimes; Yahoo Finance
Jul 3
Reports say SK hynix removed price caps from long-term memory supply agreements while Micron kept ceilings anchored to Q2 market prices — 2027 memory pricing power being contractually locked in
PCCentral; Crypto Briefing (secondary; no primary disclosure yet)
Jun 28
Korean industry reporting puts SK hynix at roughly two-thirds of NVIDIA's 2026 Vera Rubin HBM4 allocation, with Samsung expected to begin supplying NVIDIA in July after final quality tests
Maeil Business
Jul 1
Oxmiq Labs (Raja Koduri) raised $35M led by Fundomo and Samsung Catalyst Fund to license OxCore GPU IP — an 'Arm for AI silicon' play lowering the entry cost of custom accelerators
SiliconANGLE

What this means

Memory is completing its move from allocation story to capital-markets story: a ~$29.4B SEC-filed listing and reported cap-free long-term contracts mark the scarcity to market. Operators should lock HBM-backed capacity pricing before 2027 contracts reprice, and investors should watch the ~Jul 10 SKHY debut as the cleanest public-market referendum yet on the memory supercycle.

Networking.

Jun 30
PJM opened the 2028/29 capacity auction (bids close Jul 7, results Jul 14 — a week later than previously expected) and advanced a proposal for a September-October backstop procurement against the anticipated capacity shortfall
PJM Inside Lines; American Public Power Association
Jul 9 / Aug 17
FERC's large-load show-cause clock ticked into July: motions to intervene in all six RTO/ISO dockets are due Jul 9; tariff responses due Aug 17
Husch Blackwell regulatory update
Jun 24 (in-window coverage)
CoreWeave's Stockholm expansion (Conapto colocation, 32MW live) pairs Blackwell and Vera Rubin with Quantum-X800 InfiniBand — a flagship Vera Rubin deployment shipping on InfiniBand, not Spectrum-X CPO Ethernet
CoreWeave; BeBeez

What this means

No new optics event landed this week — the networking lens is power-process dominated, and the calendar now has three hard dates (Jul 9, Jul 14, Aug 17) that will reprice siting economics before any fabric decision does. Architects should note the CoreWeave data point: flagship Vera Rubin clusters can still ship on InfiniBand, so treat CPO Ethernet as an option to be priced, not an assumption.

Capital flow

Money in, revenue out.

4 categories tracked. Capital deployment up in 1 of 4; revenue follows at multiples of 0.21 to 0.6.

The four-category scorecard. Where capital is going in, where revenue is coming out, and how much of it is real. The one chart for the boardroom.

What’s real, what’s noise.

5 claims this week — 3 signal, 2 noise.

Each claim is scored 1–5 on source quality and triangulation. Anything 2 or below is flagged as noise. Where consensus is wrong, we say so.

5 / 5
SK hynix's ~$29.4B Nasdaq ADS filing marks the memory supercycle to public markets.
Sources: SEC registration statement; IBTimes; Yahoo Finance
The cleanest audited-document signal of the week. An at-range or above-range pricing around Jul 10 validates memory scarcity with public-market money rather than analyst notes; a weak debut would be the first real crack in the thesis.
4 / 5
Claude Sonnet 5's $2/$10 launch is a closed-lab price response to open-weight and efficiency pressure.
Sources: Anthropic announcement; TechCrunch
The first flagship-adjacent price cut since the open-weight cost gap widened. Watch whether OpenAI's Terra GA (promised at half GPT-5.5 cost) confirms a repricing cycle rather than a one-off — buyers should time contract renegotiations to that confirmation.
3 / 5
Meta Compute repriced the neocloud category ~15% in a morning.
Sources: Bloomberg (anon-sourced); CNBC confirmation; observable market moves
The business plan is early and could change, but the market reaction is the durable fact: neocloud equity now embeds hyperscaler-competition and customer-concentration risk it did not price on Jun 30. Investors should treat the repricing, not the plan, as the signal.
1 / 5 — noise
Grok 4.5 performs 'close to or exceeding' Claude Opus, per xAI.
Sources: Musk posts on X; TechTimes secondary coverage
Noise until proven otherwise: private beta confined to SpaceX and Tesla, no public API, no independent benchmark, and a claimed monthly from-scratch-model cadence with no documented precedent. Procurement teams should not move on this claim.
2 / 5 — noise
Meta's 'Watermelon' Muse Spark update has internally caught GPT-5.5.
Sources: Business Insider (anonymous townhall sourcing); Wang's public post promises only 'big improvements'
Internal parity claims from labs trailing the frontier have a poor verification record. The verifiable part is that a Muse Spark update with a new API is coming; wait for third-party evals before treating parity as real.

Synthesis

The week, reasoned through.

3 cross-domain connections, 5 hypotheses tested (1 under pressure), 3 patterns tracked.

Reporting says what happened; this section says what it means when you put the pieces together. Every inference is labeled by type, linked to its evidence, and held against the working framework — so when the reasoning is wrong, you can see exactly where.

Connecting the dots

Abductive
68%
confidence
Meta Compute and SB Neo are best explained as monetization responses to the hyperscaler free-cash-flow crossover — meaning the compute glut is arriving as a financing problem before it arrives as a price problem.
1. 01The Epoch analysis — top-5 hyperscaler cash capex crossing aggregate operating cash flow around Q3 2026, on ~$725B of 2026 top-4 capex — became the week's consensus macro frame.
2. 02Bloomberg reported Jul 1 that Meta is building a cloud business to sell excess AI compute; Meta rose 10% while CoreWeave and Nebius fell 12-15% on customer-concentration risk.
3. 03SoftBank announced SB Neo on Jul 2 — a Delaware neocloud targeting 10GW by ~2030 — the second hyperscale balance sheet to enter the compute-rental market in 48 hours.
Evidence: Epoch AI capex-vs-cash-flow analysis · Bloomberg: Meta building cloud business · SoftBank SB Neo press release
Deductive
71%
confidence
With the model layer cutting prices and the capacity layer adding sellers, margin is migrating to scarce physical inputs — memory and power — and the two hard-number tests of that migration land within ten days (SKHY debut ~Jul 10, PJM results Jul 14).
1. 01Anthropic cut Sonnet 5 to $2/$10 intro pricing and OpenAI promised Terra at half GPT-5.5's cost — the model layer is repricing downward.
2. 02Meta Compute and SB Neo add hyperscale sellers to compute rental, pressuring capacity margins from the supply side.
3. 03SK hynix filed a ~$29.4B Nasdaq listing and reportedly removed price caps from long-term memory contracts, while PJM capacity has cleared at the FERC cap two auctions running.
Evidence: Anthropic Claude Sonnet 5 launch · SK hynix Nasdaq filing coverage · PJM 2028/29 auction opening
Inductive
66%
confidence
The US government has become a standing variable in the frontier release pipeline: availability of the best models is now set by regulatory process at least as much as by lab readiness.
1. 01OpenAI's GPT-5.6 preview was limited to ~20 vetted partner organizations at the US government's request, with GA timing left open.
2. 02Claude Fable 5 was restored globally Jul 1 only after export controls were withdrawn; Mythos 5 remains gated to an ~100-organization 'Annex A' allowlist.
3. 03Anthropic's government-ID verification policy takes effect Jul 8 — the likely mechanism for staged, citizenship-verified frontier access.
Evidence: OpenAI GPT-5.6 preview · Anthropic Fable 5 redeployment

Thesis test

The five standing hypotheses of the working framework, tested deductively against this week’s evidence. A framework that is never strained is not being tested.

Hypothesis 1
supported
The cycle is accelerating, not slowing.
The hypothesis predicts compressing cadences across the flywheel. This week's evidence is on the release-and-pricing side: three frontier availability events in five days (Sonnet 5 GA, GPT-5.6 preview, Fable 5 restoration) and a price structure — Sonnet 5 at $2/$10, Terra promised at half GPT-5.5 — that pulls next-generation economics forward before the prior generation has amortized. Price-performance is compounding faster than deployment cycles can absorb it.
Evidence: Claude Sonnet 5 launch · GPT-5.6 preview and tiering
Hypothesis 2
supported
Capital is concentrated, returns are diffuse.
The crossover frame sharpened the concentration side: ~$725B of 2026 top-4 capex is now consuming the spenders' free cash flow, forcing Meta and SoftBank to monetize excess capacity. The diffusion side got direct evidence too — model-layer price cuts hand value to buyers, and the neocloud selloff shows the market repricing whether concentrated capital earns its return at all. The spread between where capital pools and where returns land widened this week.
Evidence: Epoch capex-vs-cash-flow analysis · Neocloud repricing on Meta Compute report
Hypothesis 3
strained
Networking is the durable layer.
No new optics or fabric event landed in-window, and the one relevant data point cuts against the strong version of the hypothesis: CoreWeave's flagship Stockholm Vera Rubin deployment ships on Quantum-X800 InfiniBand rather than the CPO Ethernet fabric NVIDIA has positioned as the AI-factory default. One deployment is not a refutation, but a framework that is never strained is not being tested — and this week strained it.
Evidence: CoreWeave Conapto Stockholm deployment
Hypothesis 4
supported
Open weights pull the floor up.
The mechanism the hypothesis describes — open-weight cost pressure forcing closed-lab repricing — visibly operated this week: Sonnet 5 launched at $2/$10 into a market where GLM-5.2 sits at ~1/6 frontier cost, and Bridgewater reported a tuned open Qwen3-235B beating its strongest frontier model on finance tasks at 13.8x lower cost. The closed price cuts are the floor pulling up, not the ceiling falling.
Evidence: Bridgewater / Thinking Machines tuned-Qwen result · Claude Sonnet 5 pricing
Hypothesis 5
supported
Power is the binding constraint for the next 24 months.
PJM opened the 2028/29 auction under an at-cap setup and simultaneously proposed a September-October backstop procurement against an anticipated capacity shortfall — the system operator itself is now planning for scarcity. Time-to-power held at 60-84 months, and the FERC docket calendar (Jul 9, Aug 17) is where siting economics get decided. Nothing this week suggested power is loosening before 2028.
Evidence: PJM 2028/29 auction and backstop proposal

Pattern watch

Inductive3 weeks observed
Grid process — not hardware supply — is the recurring headline constraint on AI buildout.
- W25: FERC issued its Jun 18 show-cause orders on large-load interconnection across six RTO/ISOs.
- W26: the orders stayed central to siting decisions; time-to-power held at 60-84 months with the 60-day tariff clock running.
- W27: PJM opened the 2028/29 auction (results Jul 14), proposed a fall backstop procurement for the expected shortfall, and the FERC intervenor deadline landed on the calendar (Jul 9).
Next week: The PJM 2028/29 result on Jul 14 clears within 5% of the ~$325 cap (formalized as prediction p54). A materially lower print would break the pattern and reopen the case that power scarcity is regional, not structural.
Inductive3 weeks observed
HBM supply is tightening generation-over-generation into a seller's market.
- W25: all three HBM makers reported volume-shipping HBM4 12-high, with allocation chatter building.
- W26: Micron confirmed 2026 HBM supply fully contracted, HBM4 ramping ~2x faster than HBM3E, and $1B+ HBM4 revenue shipped.
- W27: SK hynix filed a ~$29.4B Nasdaq listing, reportedly removed price caps from long-term contracts, and Korean reporting put it at ~two-thirds of NVIDIA's 2026 HBM4 allocation.
Next week: SK hynix's SKHY debut (~Jul 10) prices at or above range and its late-July earnings call confirms allocation exhaustion or 2027 price increases. A weak debut plus cautious earnings language would be the first counter-evidence.
Inductive3 weeks observed
US government action is now a standing gate on frontier-model availability.
- W25: Fable 5 and Mythos 5 were suspended under export controls (Jun 12).
- W26: Mythos 5 access was restored for ~100 'Annex A' critical-infrastructure organizations (Jun 26).
- W27: GPT-5.6 previewed to only ~20 government-vetted partners at the government's request; Fable 5 restored globally Jul 1 after control withdrawal.
Next week: Anthropic's government-ID verification policy ships on schedule Jul 8, and GPT-5.6's broad GA timing is visibly set by the government process rather than OpenAI's product calendar. A clean, ungated GA within two weeks would weaken the pattern.

Second-order effects

Trigger: Meta Compute and SB Neo enter compute rental.
Neocloud cost of capital rises as equity markets embed hyperscaler-competition risk — marginal neocloud buildouts get costlier financing, slowing independent capacity additions even as total sell-side capacity grows. Enterprise buyers gain a stronger negotiating position on multi-year GPU contracts than at any point since 2024.
Horizon: H1 2027Who moves: Listed neoclouds and their lenders; enterprises negotiating multi-year compute contracts
Trigger: Closed-lab price cuts (Sonnet 5 at $2/$10; Terra promised at half GPT-5.5 cost).
Agent-workload unit economics improve enough to pull forward enterprise automation deployments that were marginal at prior pricing — which generates the inference demand the new sell-side capacity needs. The price cut is simultaneously margin compression at the model layer and demand generation for the compute layer the same companies are financing.
Horizon: Q4 2026Who moves: Enterprise AI budget owners, inference clouds, and SaaS vendors defending seat pricing
Trigger: Government gating of frontier releases (GPT-5.6's vetted-partner preview; Annex A allowlists; ID verification from Jul 8).
Enterprise access to the best models becomes eligibility-dependent: organizations on government allowlists gain a structural capability edge, and multi-model routing shifts from a cost optimization to a continuity requirement. Procurement teams start writing regulatory-availability clauses into model contracts.
Horizon: Q4 2026Who moves: Enterprises outside allowlists, model-routing platform vendors, and procurement/legal teams

Strategic outlook

If you allocate capital or run infrastructure, this week reframed the 12-month posture: the scarce assets are memory and power, the crowded assets are model capability and — for the first time — raw compute capacity. Lock inference pricing while the closed-lab price war runs (Sonnet 5 intro rates end Aug 31; Terra GA is the confirmation signal), but diversify compute-vendor exposure before hyperscaler sell-side capacity resets the market's structure — anchor-customer concentration just became the neocloud category's defining risk. Treat the Jul 10 SKHY debut and the Jul 14 PJM print as the two highest-information events of the month: together they will tell you whether the physical-scarcity trade is still underpriced. And build model-availability contingency into architecture reviews now — the government is in the release pipeline, and routing flexibility has become a continuity control, not a cost lever.

Early warning panel

The levers we monitor.

10 metrics tracked — 0 rising, 2 falling, 8 steady.

Current vs prior period. Each metric has a threshold where the read materially changes — this panel flags the inflection before it lands in headlines. Click any metric for the methodology and this-week read.

Metric

Current

Prior

Dir

Threshold

Frontier lab cash position (avg months runway, top 3)
~34-37 mo; OpenAI leaning 2027 IPO, Anthropic holding Oct 2026
~34-37 mo (flat; no new in-window round)
→
<18 mo triggers re-rating risk
What this measures
Top 3 frontier labs (OpenAI, Anthropic, Google DeepMind) by disclosed runway, with xAI now public inside SpaceX. No new financing event in-window; both S-1s remain confidential-draft. New color: Bloomberg has OpenAI leaning toward a 2027 listing with a $1T valuation floor while Anthropic holds an October 2026 Nasdaq target, and Deutsche Bank frames both raises at ~$60B. Capital access stays wide; boards should not assume funding pressure forces near-term commercial concessions.
Hyperscaler capex / AI revenue ratio (top 4 weighted)
~5.0-5.3; free-cash-flow crossover framed for ~Q3 2026
~5.0-5.3 (flat; next top-4 earnings catalyst pending)
→
>6.0 invites investor pushback at next earnings
What this measures
Top 4 hyperscalers (MSFT, GOOG, META, AMZN) weighted aggregate of capex divided by AI-attributable revenue. No new top-4 print in-window, but the Epoch analysis — top-5 cash capex crossing aggregate operating cash flow around Q3 2026, on ~$725B of 2026 top-4 capex — became the consensus frame, and Meta Compute plus SB Neo look like monetization responses to it. Alphabet's Jul 28 print is the first test against actuals.
CoreWeave revenue backlog
~$100B reported; Meta Compute repriced concentration risk (stock -12-15% Jul 1)
~$100B reported / ~$131B analyst-estimated by end-Q2
→
Conversion velocity matters more than gross figure
What this measures
Booked but unrecognized revenue; the official next print is Q2 in early August. The number did not move this week but its risk profile did: Meta's reported compute-rental plan puts concentration risk on CoreWeave's ~$21B Meta contract (and Nebius's ~$27B), and the market repriced both names 12-15% in a session. Operators should watch conversion velocity and anchor-customer diversification, not the headline figure.
NVIDIA Q-over-Q data center revenue
$75.2B Q1 FY27; Q2 guide $91B, reports Aug 26
$75.2B Q1 FY27; Q2 guide $91B, reports Aug 26
→
Q2 FY27 guide $91B implies further +21% QoQ
What this measures
No within-window NVIDIA event; the next earnings print is Aug 26 against a $91B Q2 guide. In-window context supports the ramp: Korean reporting puts SK hynix at ~two-thirds of 2026 Vera Rubin HBM4 allocation with Samsung starting NVIDIA supply in July, and Vera Rubin NVL4 OEM systems remain on Q4 availability. Packaging and power, not demand, stay the binding constraints.
Open vs closed gap on coding (SWE-Bench / agentic)
Narrowing from both sides: closed prices down (Sonnet 5 $2/$10; Terra promised at half GPT-5.5), open pressure sustained
Open pressure sustained: GLM-5.2 / DeepSeek V4 Pro remain the cost challengers
↓
Sustained open lead reshapes enterprise procurement
What this measures
The gap is now compressing from the closed side too: Sonnet 5 launched at $2/$10 intro pricing with 63.2% agentic coding (vs Opus 4.8's 69.2%), and GPT-5.6 Terra is promised at half GPT-5.5's cost. On the open side, Bridgewater and Thinking Machines reported a tuned Qwen3-235B beating their strongest frontier model on internal finance tasks at 13.8x lower cost (company-run eval). Architects should re-run cost-per-task benchmarks this month — the frontier just got cheaper in both directions.
Sovereign AI commitments (count / aggregate $)
~14 / ~$180B+ (flat; SB Neo's 10GW is corporate, not sovereign)
~14 / ~$180B+ (flat; no new drawn sovereign mega-commitment)
→
—
What this measures
Analyst-curated count of sovereign/national AI-compute commitments. No new drawn commitment in-window. SB Neo's 10GW US target (launching FY2027 on SB Energy's DOE-land Ohio campus) is corporate capital on state-adjacent infrastructure, and the reported ~$590B Korean chip-hub plan is semiconductor manufacturing, not compute commitment — and still needs primary confirmation. Operators should not count either toward sovereign compute capacity yet.
PJM 2026/27 capacity auction price ($/MW-day)
$329.17; 2028/29 BRA bids close Jul 7, results Jul 14 (slipped ~1 week)
$329.17; 2028/29 BRA results expected around Jul 7
→
11x in 24 months — power is the new binding constraint
What this measures
The 2026/27 BRA cleared at the FERC cap ($329.17); 2027/28 cleared at $333.44. The 2028/29 auction opened Jun 30 with bids closing Jul 7 and results now posting Jul 14 after 4 p.m. ET (cap ~$325, floor $175) — a one-week slip from prior expectations. PJM also proposed a September-October backstop procurement against the anticipated capacity shortfall. Architects should budget capacity at-cap through 2028 and treat a sub-cap print as the upside surprise, not the base case.
Time-to-power, busiest US markets (months)
60-84; FERC intervenor deadline Jul 9, tariff responses due Aug 17
60-84; FERC 60-day tariff-response clock is the next catalyst
→
—
What this measures
Months from new-load interconnection request to energization. The FERC show-cause process hit its procedural stride: motions to intervene in all six RTO/ISO dockets are due Jul 9, and tariff responses or revisions are due Aug 17. Large-power-transformer lead times remain ~128 weeks. No near-term change to the 60-84 month reality — architects should still pre-commit power and long-lead grid equipment before GPU SKUs.
Cost-per-task, frontier reasoning model
DeepSeek V4 official (mid-July) adds peak/off-peak surge pricing — the first major time-of-day price discrimination
~$0.08-$0.13 effective; Sonnet 5 $2/$10 intro, GPT-5.6 tiering Sol $5/$30 / Terra $2.50/$15 / Luna $1/$6
~$0.10-$0.15 effective; inference-cloud competition rising
↓
—
What this measures
Median cost across frontier-tier reasoning models for a benchmark complex task. The ceiling dropped in-window: Sonnet 5 at $2/$10 intro pricing, GPT-5.6's three-tier structure topping out at $5/$30, and Sol landing on Cerebras at up to 750 tok/s in July. One caveat from the Pulse: Sonnet 5's higher token appetite means sticker price and cost-per-completed-task have diverged — operators should benchmark on completed tasks, not rate cards.
Custom silicon share of incremental AI compute
~33-36%; silicon-IP layer forming (Oxmiq) to lower custom-ASIC entry cost
~33-36%; HBM and CPO now more binding than raw accelerator demand
→
>35% materially compresses merchant GPU pricing
What this measures
The J.P. Morgan framing (2026 custom-ASIC market ~$60-70B, Broadcom 80-85%, Marvell 10-12%) still stands. This week's supporting data point: Oxmiq Labs raised $35M (Samsung Catalyst Fund, MediaTek, Pegatron participating) to license GPU IP that lowers the design cost of custom accelerators — an 'Arm for AI silicon' layer that would accelerate the ASIC-share thesis if it lands customers. Investors should keep diversifying into the co-design duopoly plus packaging and power.

Frontier lab cash position (avg months runway, top 3)
~34-37 mo; OpenAI leaning 2027 IPO, Anthropic holding Oct 2026→vs ~34-37 mo (flat; no new in-window round)
Threshold: <18 mo triggers re-rating risk
What this measures
Top 3 frontier labs (OpenAI, Anthropic, Google DeepMind) by disclosed runway, with xAI now public inside SpaceX. No new financing event in-window; both S-1s remain confidential-draft. New color: Bloomberg has OpenAI leaning toward a 2027 listing with a $1T valuation floor while Anthropic holds an October 2026 Nasdaq target, and Deutsche Bank frames both raises at ~$60B. Capital access stays wide; boards should not assume funding pressure forces near-term commercial concessions.
Hyperscaler capex / AI revenue ratio (top 4 weighted)
~5.0-5.3; free-cash-flow crossover framed for ~Q3 2026→vs ~5.0-5.3 (flat; next top-4 earnings catalyst pending)
Threshold: >6.0 invites investor pushback at next earnings
What this measures
Top 4 hyperscalers (MSFT, GOOG, META, AMZN) weighted aggregate of capex divided by AI-attributable revenue. No new top-4 print in-window, but the Epoch analysis — top-5 cash capex crossing aggregate operating cash flow around Q3 2026, on ~$725B of 2026 top-4 capex — became the consensus frame, and Meta Compute plus SB Neo look like monetization responses to it. Alphabet's Jul 28 print is the first test against actuals.
CoreWeave revenue backlog
~$100B reported; Meta Compute repriced concentration risk (stock -12-15% Jul 1)→vs ~$100B reported / ~$131B analyst-estimated by end-Q2
Threshold: Conversion velocity matters more than gross figure
What this measures
Booked but unrecognized revenue; the official next print is Q2 in early August. The number did not move this week but its risk profile did: Meta's reported compute-rental plan puts concentration risk on CoreWeave's ~$21B Meta contract (and Nebius's ~$27B), and the market repriced both names 12-15% in a session. Operators should watch conversion velocity and anchor-customer diversification, not the headline figure.
NVIDIA Q-over-Q data center revenue
$75.2B Q1 FY27; Q2 guide $91B, reports Aug 26→vs $75.2B Q1 FY27; Q2 guide $91B, reports Aug 26
Threshold: Q2 FY27 guide $91B implies further +21% QoQ
What this measures
No within-window NVIDIA event; the next earnings print is Aug 26 against a $91B Q2 guide. In-window context supports the ramp: Korean reporting puts SK hynix at ~two-thirds of 2026 Vera Rubin HBM4 allocation with Samsung starting NVIDIA supply in July, and Vera Rubin NVL4 OEM systems remain on Q4 availability. Packaging and power, not demand, stay the binding constraints.
Open vs closed gap on coding (SWE-Bench / agentic)
Narrowing from both sides: closed prices down (Sonnet 5 $2/$10; Terra promised at half GPT-5.5), open pressure sustained↓vs Open pressure sustained: GLM-5.2 / DeepSeek V4 Pro remain the cost challengers
Threshold: Sustained open lead reshapes enterprise procurement
What this measures
The gap is now compressing from the closed side too: Sonnet 5 launched at $2/$10 intro pricing with 63.2% agentic coding (vs Opus 4.8's 69.2%), and GPT-5.6 Terra is promised at half GPT-5.5's cost. On the open side, Bridgewater and Thinking Machines reported a tuned Qwen3-235B beating their strongest frontier model on internal finance tasks at 13.8x lower cost (company-run eval). Architects should re-run cost-per-task benchmarks this month — the frontier just got cheaper in both directions.
Sovereign AI commitments (count / aggregate $)
~14 / ~$180B+ (flat; SB Neo's 10GW is corporate, not sovereign)→vs ~14 / ~$180B+ (flat; no new drawn sovereign mega-commitment)
What this measures
Analyst-curated count of sovereign/national AI-compute commitments. No new drawn commitment in-window. SB Neo's 10GW US target (launching FY2027 on SB Energy's DOE-land Ohio campus) is corporate capital on state-adjacent infrastructure, and the reported ~$590B Korean chip-hub plan is semiconductor manufacturing, not compute commitment — and still needs primary confirmation. Operators should not count either toward sovereign compute capacity yet.
PJM 2026/27 capacity auction price ($/MW-day)
$329.17; 2028/29 BRA bids close Jul 7, results Jul 14 (slipped ~1 week)→vs $329.17; 2028/29 BRA results expected around Jul 7
Threshold: 11x in 24 months — power is the new binding constraint
What this measures
The 2026/27 BRA cleared at the FERC cap ($329.17); 2027/28 cleared at $333.44. The 2028/29 auction opened Jun 30 with bids closing Jul 7 and results now posting Jul 14 after 4 p.m. ET (cap ~$325, floor $175) — a one-week slip from prior expectations. PJM also proposed a September-October backstop procurement against the anticipated capacity shortfall. Architects should budget capacity at-cap through 2028 and treat a sub-cap print as the upside surprise, not the base case.
Time-to-power, busiest US markets (months)
60-84; FERC intervenor deadline Jul 9, tariff responses due Aug 17→vs 60-84; FERC 60-day tariff-response clock is the next catalyst
What this measures
Months from new-load interconnection request to energization. The FERC show-cause process hit its procedural stride: motions to intervene in all six RTO/ISO dockets are due Jul 9, and tariff responses or revisions are due Aug 17. Large-power-transformer lead times remain ~128 weeks. No near-term change to the 60-84 month reality — architects should still pre-commit power and long-lead grid equipment before GPU SKUs.
Cost-per-task, frontier reasoning model
~$0.08-$0.13 effective; Sonnet 5 $2/$10 intro, GPT-5.6 tiering Sol $5/$30 / Terra $2.50/$15 / Luna $1/$6↓vs ~$0.10-$0.15 effective; inference-cloud competition rising
DeepSeek V4 official (mid-July) adds peak/off-peak surge pricing — the first major time-of-day price discrimination
What this measures
Median cost across frontier-tier reasoning models for a benchmark complex task. The ceiling dropped in-window: Sonnet 5 at $2/$10 intro pricing, GPT-5.6's three-tier structure topping out at $5/$30, and Sol landing on Cerebras at up to 750 tok/s in July. One caveat from the Pulse: Sonnet 5's higher token appetite means sticker price and cost-per-completed-task have diverged — operators should benchmark on completed tasks, not rate cards.
Custom silicon share of incremental AI compute
~33-36%; silicon-IP layer forming (Oxmiq) to lower custom-ASIC entry cost→vs ~33-36%; HBM and CPO now more binding than raw accelerator demand
Threshold: >35% materially compresses merchant GPU pricing
What this measures
The J.P. Morgan framing (2026 custom-ASIC market ~$60-70B, Broadcom 80-85%, Marvell 10-12%) still stands. This week's supporting data point: Oxmiq Labs raised $35M (Samsung Catalyst Fund, MediaTek, Pegatron participating) to license GPU IP that lowers the design cost of custom accelerators — an 'Arm for AI silicon' layer that would accelerate the ASIC-share thesis if it lands customers. Investors should keep diversifying into the co-design duopoly plus packaging and power.

Predictions

What we expect next.

4 predictions for the next 30-90 days, confidence 62%-74%.

Each prediction is falsifiable, time-bounded, and tied to a specific signal we will watch. Future issues score these hit, miss, partial, or pending and build a public track record.

Prediction 01

67%

confidence

Capital

SK hynix's Nasdaq ADS offering prices at or above its indicated ~$166/ADS level and closes its first trading week above the offer price, by July 31, 2026.

Deadline: By July 31, 2026

Trigger: SKHY pricing announcement and first-week Nasdaq trading data (~Jul 10 debut expected).

Prediction 02

74%

confidence

Power

The PJM 2028/29 base residual auction clears within 5% of the ~$325/MW-day cap when results post on July 14, 2026.

Deadline: By July 14, 2026

Trigger: PJM BRA results publication (after 4 p.m. ET, Jul 14).

Prediction 03

62%

confidence

Software

GPT-5.6 reaches broad GA with the Terra tier priced at or below $2.50/$15 per MTok — half of GPT-5.5's rate — confirming a closed-lab repricing cycle rather than a one-off Sonnet 5 cut, by August 31, 2026.

Deadline: By August 31, 2026

Trigger: OpenAI pricing page / API changelog at GPT-5.6 general availability.

Prediction 04

66%

confidence

Hardware

Samsung's HBM4 supply to NVIDIA is publicly confirmed — via earnings call, company statement, or multi-source supply-chain reporting — by August 31, 2026.

Deadline: By August 31, 2026

Trigger: Samsung Q2 earnings call (late July), NVIDIA disclosure, or corroborated supply-chain reporting.

Track record

Scoring prior predictions.

4 prior predictions: 2 hit, 0 miss, 0 partial, 2 pending. Hit rate 100%.

4 predictions across issues so far. Hit rate: 100%. Hits 2, misses 0, partials 0, pending 2.

Prediction 01

72%

confidence

Hardware

By July 31, 2026, at least one additional memory supplier besides Micron publicly confirms 2026 HBM4 supply is fully allocated or materially price-up for 2027.

Deadline: By July 31, 2026

Trigger: Samsung or SK hynix earnings call, investor presentation, or supply-chain report.

hit— Hit. Samsung's CFO stated at the Apr 30 call that 2026 HBM4 output is fully sold out, and in-window reporting says SK hynix removed price caps from long-term agreements — explicit 2027 price-up posture. Caveat: the in-window items are supply-chain reporting; SK hynix's late-July earnings call would make it unambiguous.

Prediction 02

57%

confidence

Capital

Groq announces at least one named Fortune 500 or hyperscaler inference-cloud customer by September 30, 2026.

Deadline: By September 30, 2026

Trigger: Groq customer announcement, case study, or partner release.

pending— Pending, no in-window evidence. Slightly harder now: Meta Compute and SB Neo add competition for exactly the enterprise inference customers Groq needs to name.

Prediction 03

63%

confidence

Networking

A named Vera Rubin partner announces a CPO/Spectrum-X Ethernet Photonics rack or cluster design win by August 31, 2026.

Deadline: By August 31, 2026

Trigger: NVIDIA, OEM, cloud, or networking vendor product/customer announcement.

pending— Pending; this week's data point cuts against — CoreWeave's Stockholm Vera Rubin deployment ships on Quantum-X800 InfiniBand, not CPO Ethernet. The GTC Taipei adopter list (CoreWeave, Lambda, OCI) still provides the setup.

Prediction 04

66%

confidence

Software

At least one major enterprise platform ships an admin control specifically for scheduled/background coding or app-building agents by August 31, 2026.

Deadline: By August 31, 2026

Trigger: Product changelog or GA announcement from OpenAI, Cursor, Microsoft, ServiceNow, GitHub, or Atlassian.

hit— Hit. GitHub shipped Copilot agent session streaming to public preview (Jul 2) — SIEM/Purview streaming of all agent sessions — on top of its agent control plane, and GitHub also added AI-credit session limits covering background agents (Jul 1, per Agent Techniques coverage).

Watchlist

On the radar this week.

5 catalysts to watch, starting Jul 7-9.

Specific catalysts that would change the read materially. Watching these tells us whether the thesis is strengthening or weakening.

Jul 7-9
PJM bids close (Jul 7), Fable 5 usage-credit pricing begins (Jul 7), FERC intervenor deadline (Jul 9)
Three hard dates in 72 hours: the auction sets up the Jul 14 print, Fable 5's credit transition is a live test of post-restoration Claude demand, and the FERC intervenor list will show who intends to fight over large-load tariffs.
~Jul 10
SK hynix SKHY ADS begins trading on Nasdaq
Pricing versus the ~$166/ADS, ~$29.4B expectation is a public-market referendum on the memory supercycle — the week's cleanest test of the scarcity thesis with real money.
Jul 14
PJM 2028/29 capacity auction results (after 4 p.m. ET)
An at-cap clearing confirms power as the binding AI-factory constraint through 2028; a materially sub-cap print would be the first crack in the pattern and would reprice siting strategy.
Mid-late Jul
GPT-5.6 broad GA, DeepSeek V4 official (with peak/off-peak pricing), Gemini 3.5 Pro July target
Three release events that will decide whether the Sonnet 5 price cut was a one-off or the start of a closed-lab repricing cycle — the single biggest input to H2 inference budgeting.
Jul 28
Alphabet Q2 earnings — first top-4 hyperscaler print of the season
First test of the free-cash-flow-crossover narrative against actuals; capex guidance and AI-revenue disclosure will recalibrate the capex-to-revenue lever for the whole category.

Companion reads

The rest of the spine.

The AI Stack Weekly is the cross-stack flywheel read. Pair it with the model-and-tree spine and the working framework to get the full picture.