Daily Investment Brief — September 11, 2026

Market Pulse

The compute pricing market is undergoing a clean structural bifurcation between inference and training hardware — not a broad repricing, but a divergence between two distinct demand regimes. Inference-class ASICs and GPUs (AWS Inferentia SPOT +131–484% across US regions over 30 days; NVIDIA L4 ON_DEMAND +98% in Virginia; VIRTEX_VU47P FPGA SPOT +112–248% in Virginia and London) are running into hard supply ceilings as real-time inference workloads crowd out idle capacity. Training-class ASICs are moving in the opposite direction: GCP TPU V5 LitePod SPOT collapsed -75% in Virginia and -89% in Singapore over the same 30-day window, while AWS Trainium1 SPOT in Oregon fell -63% to $0.079/hr — a 92% discount to on-demand list. The dominant narrative is generational hardware transition: AWS and GCP are simultaneously flooding the market with surplus gen-1 training capacity while demand piles into inference-optimized silicon. On-demand and reserved prices for both Trainium and Inferentia are completely frozen (0.0% change over 30 days), creating an unusually wide and stable gap between list prices and actual market-clearing SPOT rates that is the single most actionable pricing signal in today's dataset.

Key Movers

Noise alert: UNMAPPED vCPU/RAM tickers in Ohio, Seoul, Stockholm, and Virginia are showing multi-million-percent 24h/30d deltas. These are un-normalized Oracle/niche-provider entries and should be disregarded entirely. Brazil Trainium SPOT ($4.06/hr) is a thin-market artifact, not real inference demand.

Investable Insights

H1 — AWS Inferentia Demand Recovery Points to Structural Inference Tightening and Gen3 Catalyst Confidence: 4 / 5

Thesis: AWS Inferentia SPOT prices across all four US regions (Ohio +484%, Oregon +133%, Virginia +131%, Mumbai +60% over 30 days) are recovering from an extreme August trough — not spiking from a stable base. The Ohio per-chip SPOT price fell from $0.034/hr in January to a floor of $0.000085/hr around August 10, then climbed back to $0.012/hr today. That nadir was peak surplus: AWS was flooding the SPOT pool during a lower-demand window. The current recovery means the inference ASIC SPOT pool has been fully absorbed by genuine workload demand — both inf1.24xlarge (16× gen-1 chips, $0.52–$0.67/hr range in Ohio) and inf2.48xlarge (12× Inferentia2 chips, $1.54–$1.73/hr) are now priced with near-zero intraday variance, meaning these instance families have cleared the market and AWS has no idle surplus left to dump. The structural implication is durable: Inferentia hardware is now 6.8 years into its lifecycle (launched December 2019, still listed is_active: true on AWS only), there has been a complete public information blackout on Inferentia roadmap for months, and AWS re:Invent is expected in November 2026. This is the textbook setup for a generational product transition. On-demand pricing has not moved (0.0% over 30 days), confirming AWS is not cutting to clear — it is holding price while SPOT clears naturally.

Key Evidence:

1. Ohio Inferentia SPOT recovered from a trough of $0.000085/hr (August 10) to $0.012/hr today — a +484% 30d move that is a recovery from surplus, not a new supply squeeze. (Live ticker data, September 10, 2026)
2. Both inf1.24xlarge and inf2.48xlarge coexist in Ohio SPOT pool simultaneously, with stable tick-to-tick pricing and no intraday auction noise — full market clearance confirmed. (Underlying instance drill-down, September 10, 2026)
3. Inferentia on-demand pricing: 0.0% change over 30 days across all US regions — AWS is not discounting to clear, consistent with a stable list-price anchor ahead of a new generation. (Ticker data, September 10, 2026)
4. Inferentia (launched December 2019) is the oldest active ASIC in AWS's catalog at 6.8 years, with no competitive equivalent in any other cloud provider's lineup. (Depreciation analysis)
5. Zero public communications from AWS on Inferentia roadmap — press search returned no results. Information blackout of this duration from a product team typically precedes a major announcement. (News feed, September 10, 2026)
6. Qualifying: The current $0.012/hr per-chip Ohio SPOT remains below the January level of $0.034/hr — this is still a recovery, not an overheating signal. The spot discount (currently 60–87% vs. OD across regions) is still large, meaning the market has not inverted.

Implied Action:

• Immediate: Inferentia workload operators should lock Inferentia RESERVED_1YR contracts now, before any Inferentia3 announcement triggers an on-demand price reset. Current OD pricing is the base; RI holders will retain current-generation cost structure for 12 months regardless of what Inferentia3 costs.
• Monitor trigger: Watch for inf3.* instance type appearing in the AWS pricing API — this is the clearest possible confirmation. Secondary trigger: Inferentia SPOT discount narrowing from ~80% toward <50%, which would indicate AWS is withdrawing gen-1 from the SPOT pool.
• Equity angle: AWS's custom silicon strategy ($25B+ annualized inference ASIC revenue per news coverage, September 2026) validates an AWS-specific margin expansion thesis within AMZN (currently $251.89, analyst consensus target $328). The Inferentia lifecycle event is a 60–90 day catalyst not yet discussed in sell-side notes.

H2 — GCP TPU V5 LitePod Fire Sale: Discrete Pricing Tiers Create an Exploitable Training Arbitrage Window Confidence: 3.5 / 5

Thesis: GCP TPU V5 LitePod SPOT prices are not responding to real-time auction dynamics — they are moving in step-function tiers that reflect GCP's Dynamic Workload Scheduler (DWS) and committed discount tables being revised at discrete intervals. In Virginia, the price held at $0.12713/hr for weeks, then stepped to $0.13884/hr, then fell to $0.09443/hr on August 25, then abruptly to $0.039/hr on August 27, and currently sits at $0.023/hr. Singapore shows an identical staircase pattern: $0.129 → $0.115 → $0.090 → $0.036 → $0.007 → $0.003/hr, each level held for approximately two weeks before stepping down. This is GCP policy pricing, not market clearing — and it means the current floor tier is actionable and stable for approximately 2–4 weeks before the next adjustment. At $0.003/hr in Singapore and $0.023/hr in Virginia, TPU V5 LitePod SPOT represents the cheapest large-scale AI training compute available anywhere in the public cloud. At this price, even with 15–20% utilization overhead from checkpoint/restart cycles on preemptible capacity, the effective compute cost is still 5–10× cheaper than Trainium SPOT in Virginia ($0.324/hr) or H100 ON_DEMAND. Google Cloud CEO Thomas Kurian's explicit statement at Goldman Sachs' Communacopia conference (September 10, 2026) that the TPU business is "more than twice the size of the next-largest hyperscaler's" confirms this is aggressive ecosystem land-grab pricing, not distress.

Key Evidence:

1. Virginia TPU V5 LitePod SPOT: $0.023/hr current vs. $0.094/hr 30 days ago (-75%); discrete staircase steps visible in 90-day history, not continuous auction noise. (Ticker history, September 11, 2026)
2. Singapore TPU V5 LitePod SPOT: $0.003/hr current vs. $0.031/hr 30 days ago (-89%); clean six-step staircase pattern over 90 days confirms systematic policy repricing. (Ticker history, September 11, 2026)
3. Amsterdam TPU V5 LitePod SPOT appeared from near-zero at $0.026/hr (+2,486% 7d) — GCP seeding new geographies with cheap capacity, consistent with ecosystem expansion strategy. (Top movers scan, September 10, 2026)
4. Google Cloud CEO Thomas Kurian: "TPU business is more than twice the size of the next-largest hyperscaler's" — Goldman Sachs Communacopia conference, September 10, 2026. (News feed, September 10, 2026)
5. TPU V6E and V7 tickers confirmed present in the component catalog — V5 LitePod is being deprecated in favor of next-generation hardware, consistent with fire-sale dynamics. (Component list, September 10, 2026)
6. Qualifying: GCP SPOT pricing is not a real-time auction (unlike AWS SPOT). Preemption frequency may be elevated in oversupplied markets, and each step-down tier could reset on a 2-week schedule — the $0.003/hr floor may not be the final step.

Implied Action:

• Immediate: ML teams running JAX/XLA training workloads should deploy into Singapore and Virginia TPU V5 LitePod SPOT now. At $0.003–$0.023/hr per chip, budget-constrained labs can run training runs at a fraction of H100 or even Trainium costs. Treat the current tier as a 2–4 week window.
• Operational requirement: Implement robust checkpointing with ≤15-minute intervals. The breakeven for checkpoint overhead at $0.023/hr is trivially low — even a 20% utilization tax still beats alternatives.
• Equity angle: GOOGL at $332.60 (-18.6% from May peak at $408) is being penalized for $205B capex commitments while the TPU business simultaneously claims market leadership. This is the same land-grab pricing strategy AWS used in 2013. GOOGL at current levels with a consensus target of $428 (+29%) and a 22-year Finnish nuclear power contract providing baseload stability represents a credible 18-month long. The capex fear is real but peak; the TPU ecosystem lock-in is underweighted.

H3 — Trainium Bifurcation: Oregon Gen-1 Glut Is Structural, 1YR Reserved Contracts Are a Value Trap Confidence: 4.5 / 5

Thesis: Trainium SPOT pricing has sorted cleanly into two geographically distinct regimes that reflect a confirmed generational transition in progress. Virginia SPOT is up +27.3% over 30 days to $0.324/hr and still climbing (+3.9% in 7 days) — AWS is absorbing all gen-2+ Trainium capacity in core eastern regions for enterprise reservations, leaving zero surplus. Oregon SPOT has collapsed -63.3% to $0.079/hr, representing an 80.6% discount to the $0.408/hr on-demand price — the deepest SPOT discount in any US Trainium market. The mechanism is straightforward: AWS built Trainium1 fleet density in Oregon during its initial launch phase, and as Trainium2/3/4 take center stage in Virginia and Ohio, AWS has no incentive to retire the Oregon fleet — it simply lets SPOT pricing discover the clearing price for gen-1 hardware. The commercial confirmation is decisive: an August 27, 2026 report (Stocktwits/MarketWise) states that Trainium2 is "largely sold out," Trainium3 is "nearly fully subscribed," and Trainium4 is "already reserved" 18 months before availability. Meanwhile, Trainium1 on-demand and 1YR reserved prices are completely frozen — $0.937–$0.941/hr OD and $0.617–$0.620/hr RI across all US regions, 0.0% change over 30 days. A 1YR Trainium1 RI at $0.617–$0.620/hr is the most clearly identified value trap in today's dataset: you are paying full list for gen-1 hardware that the SPOT market is clearing at 8–20 cents.

Key Evidence:

1. Oregon Trainium SPOT: $0.079/hr today vs. $0.215/hr 30 days ago (-63.3%); 80.6% discount to $0.408/hr on-demand; deepest SPOT discount in any US Trainium market. (Ticker data, September 10, 2026)
2. Virginia Trainium SPOT: $0.324/hr today, +27.3% over 30 days, +3.9% in 7 days — rising while Oregon falls. Clean bifurcation. (Ticker data, September 10, 2026)
3. Trainium1 ON_DEMAND: $0.937–$0.941/hr across all US regions; 0.0% change over 30 days. Trainium1 RESERVED_1YR: $0.617–$0.620/hr; 0.0% change over 30 days. Frozen pricing = no list-price support for gen-1 hardware. (Ticker data, September 10, 2026)
4. Trainium2 "largely sold out," Trainium3 "nearly fully subscribed," Trainium4 "already reserved" 18 months early — confirmed commercial pipeline that explains why gen-2+ capacity is unavailable in SPOT pool. (Stocktwits/MarketWise, August 27, 2026)
5. Amazon Trainium custom chip business crossed $25B+ annualized run rate — validates the commercial scale of the ASIC transition and confirms this is not a marginal product line. (News coverage, August/September 2026)
6. Brazil Trainium SPOT: $4.06/hr, -41% over 30d, -45% in 7d — artifact of thin EM SPOT pool, not a real market signal; disregard. (Ticker data — flagged as noise)

Implied Action:

• Buy: Oregon Trainium1 SPOT aggressively for batch ML training workloads. At $0.079/hr (80.6% discount to list), this is the cheapest large-ASIC training compute in the US. Checkpoint every 10 minutes to manage preemption risk. Break-even for 20% utilization overhead leaves effective cost at ~$0.099/hr — still 88% below OD.
• Avoid: Do not sign Trainium1 RESERVED_1YR contracts at $0.617/hr. The SPOT market is pricing this hardware at $0.079–$0.177/hr. Locking in $0.617/hr for 12 months on gen-1 hardware while Trainium2/3 ramp is the worst risk-adjusted position in the dataset.
• Equity angle: The velocity of AWS ASIC generational iteration (T1 → T2 → T3 → T4 in rapid succession) is the strongest pricing-data-grounded evidence of NVDA displacement in cloud training markets. AMZN's AWS margin expansion thesis gains another data point here. The mirror implication: pure-play Nvidia GPU cloud neoclouds face a structural margin compression cycle as large training labs route gen-1 ASIC budget to Oregon SPOT instead of GPU ON_DEMAND.

H4 — T4 Madrid: Market Inception Event, Not Price Spike — Reserved Contracts Offer Value at Below-NatEx Pricing Confidence: 2.5 / 5

Thesis: The NVIDIA T4's +560–668% price move in Madrid over 30 days is not a transient spike — the market structure underneath it (2 providers on ON_DEMAND, 1 SPOT provider, 330 NatEx matched pairs, a tight NatEx IQR of $0.35/$0.35 on p25/p75) is consistent with a freshly deployed regional fleet, not a pricing error. The current ON_DEMAND price of $0.450/hr sits 28.6% above the NatEx-validated fair value of $0.350/hr — a premium that is typical of thin markets where providers test price discovery before competition arrives. Crucially, reserved contracts are already forming: RESERVED_1YR is available at $0.221/hr and RESERVED_3YR at $0.150/hr, both with n=1 observations. The 1YR RI at $0.221/hr sits 37% below the NatEx fair value of $0.350/hr, meaning even in the absence of additional market participants or any price discovery dynamics, the 1YR RI is already priced at a structural discount to the model-validated equilibrium. The EU AI Act compliance narrative (workloads requiring EU-sovereign compute) and the emergence of Spain as a Southern European compute hub add medium-term demand support, but the confidence discount reflects the absence of a ticker history (the market is too young to have 90-day data) and no corroborating infrastructure news from Spain.

Key Evidence:

1. T4 ON_DEMAND Madrid: $0.450/hr today vs. near-zero 30 days ago (+560%). 2 providers confirmed on ON_DEMAND. (Ticker data, September 10, 2026)
2. NatEx validation: natex_median = $0.350/hr with tight IQR ($0.35/$0.35 on p25/p75), n_pairs = 330 — robust cross-validation despite market youth. ON_DEMAND at $0.450/hr is 28.6% above fair value. (Ticker data, September 10, 2026)
3. RESERVED_1YR at $0.221/hr — 37% below NatEx fair value of $0.350/hr; a structurally cheap commitment even assuming zero market maturation. (Ticker data, September 10, 2026)
4. No corroborating infrastructure news for Spain datacenter investment returned from news or intel feeds — a gap that limits confidence in the market-formation thesis. (News/Intel feeds, September 10, 2026)
5. No ticker price history available — market is too young to trend-validate. (Ticker history query, September 11, 2026)
6. Qualifying: n_providers = 1 on SPOT; a single provider exit would collapse the SPOT market. Concentration risk is elevated. Madrid remains an emerging market with all associated liquidity risks.

Implied Action:

• Opportunistic buy: Lock T4 RESERVED_1YR in Madrid at $0.221/hr for workloads with confirmed EU-sovereign compliance requirements. The NatEx discount alone justifies this trade on value grounds, independent of any market maturation upside.
• Monitor trigger: A second provider appearing in the Madrid T4 SPOT market within 60 days is the primary confirmation signal — it indicates competition is arriving and the 28.6% ON_DEMAND premium is about to compress. Check the raw instance table monthly.
• Position size: Keep this a small/exploratory position. The 2.5/5 confidence rating reflects genuine data limitations, not a bear case — but the risk of a single-provider market collapsing is non-trivial.

H5 — Custom Silicon Cambrian Explosion: The Inference/Training ASIC Bifurcation Is Mispriced Across the Equity Stack Confidence: 4 / 5

Thesis: The compute pricing data tells a story that the equity market is not yet fully reading: inference-class and training-class custom ASICs are operating in anti-correlated demand regimes simultaneously. Inference ASICs (Inferentia SPOT +131–484%, VIRTEX FPGA SPOT +112–248%) are supply-constrained and repricing upward while hyperscalers consume their own capacity. Training ASICs (TPU V5 LitePod SPOT -75% to -89%, Trainium1 Oregon SPOT -63%) are experiencing deliberate supply floods as generational transitions occur. The equity market is pricing this as a single "AI infrastructure" story — buy NVDA, buy hyperscalers, hold — missing the disaggregated signal. Broadcom (AVGO), the clearest custom ASIC proxy with dual Google TPU and Meta MTIA exposure, trades at 17.7× forward P/E — the lowest multiple in the AI infrastructure stack — despite sitting at the intersection of both regimes. Amazon's Trainium commercialization ($25B+ annualized run rate) and Google's explicit TPU leadership claim (CEO Kurian at Goldman Sachs Communacopia, September 10, 2026) confirm that two hyperscalers have simultaneously crossed meaningful ASIC revenue thresholds. The DOJ antitrust investigation into Nvidia's $17B Groq licensing deal adds a tail risk to the Nvidia-orchestrated "NVLink Fusion" custom silicon ecosystem — and paradoxically, ecosystem fragmentation would accelerate the multi-vendor ASIC buildout that AVGO is positioned to benefit from.

Key Evidence:

1. Inferentia SPOT +131–484% across US regions over 30 days; VIRTEX_VU47P FPGA SPOT +112–248% in Virginia and London — inference ASIC demand exceeding supply. (Ticker data, September 10, 2026)
2. TPU V5 LitePod SPOT -75% in Virginia, -89% in Singapore over 30 days; Trainium1 Oregon SPOT -63% over 30 days — training ASIC supply flood. (Ticker data, September 10, 2026)
3. Google Cloud CEO Kurian: "TPU business is more than twice the size of the next-largest hyperscaler's" — Goldman Sachs Communacopia, September 10, 2026. (News feed)
4. Amazon Trainium at $25B+ annualized run rate; Trainium2 sold out, Trainium3 nearly subscribed, Trainium4 already reserved — hyperscaler ASIC commercialization crossing inflection. (Stocktwits/MarketWise, August 27, 2026; News coverage, September 2026)
5. AVGO at $506, 17.7× forward P/E — lowest multiple in AI infrastructure stack — with dual Google TPU and Meta MTIA ASIC exposure. (Equity data, September 10, 2026)
6. DOJ investigation into Nvidia's $17B Groq licensing deal confirmed — NVLink Fusion ecosystem fragmentation risk is non-trivial and could accelerate multi-vendor ASIC adoption. (News feed, September 10, 2026)
7. MediaTek 44% YoY sales growth driven by Google ASIC production ramp to mass production Q4 2026 — supply chain confirmation that custom silicon is scaling. (News feed, September 10, 2026)
8. Qualifying: Nvidia Q2 revenue of $96.2B (+106% YoY) with Q3 guidance ~$108B — Nvidia is not losing today. The bifurcation is a medium-term trend, not a near-term Nvidia collapse thesis.

Implied Action:

• Long AVGO into the Q3 earnings print: Custom ASIC revenue trajectory for Google TPU and Meta MTIA should reflect the inference demand surge directly. At 17.7× forward P/E, this is a cheap entry into the non-Nvidia custom silicon story. Primary catalyst: Q3 earnings ASIC/XPU revenue growth rate.
• Monitor QCOM: Qualcomm's AI accelerator disclosures in the next 10-Q will either confirm or deny whether the custom silicon buildout is truly multi-vendor or still AWS/Google duopoly.
• Cautious on GPU-only neoclouds: Inference GPU SPOT tightening (Inferentia, L4) signals AWS and GCP are consuming their own custom inference silicon before selling spot capacity to third parties. Neoclouds relying on GPU SPOT arbitrage for inference revenue face margin compression from both ends.
• Watch NVDA DOJ: If the antitrust investigation into the Groq deal fractures the NVLink Fusion ecosystem, the timeline for multi-vendor ASIC stacks accelerates — a tail-risk catalyst that is currently unpriced.

Risk Flags

Immediate (0–30 days)

R1 — Trainium1 Reserved Contract Lock-in Risk The frozen Trainium1 RI price ($0.617/hr for 1YR) contrasted with Oregon SPOT at $0.079/hr creates a dangerous pricing illusion: the list price looks "safe" and the SPOT discount looks anomalous. In reality, the SPOT market is clearing at the true hardware value while RI pricing reflects AWS's contractual inertia. Any organization evaluating Trainium1 capacity expansion in the next 30 days should treat RI contracts as disqualified. The SPOT discount (80.6% in Oregon) has widened every week for 30 days with no reversal signal. Watch: if on-demand prices are cut (currently locked at 0.0%), that triggers the final confirmation that gen-1 is being deprecated — at that point, all existing RI holders take an immediate mark-to-market loss on their committed capacity.

R2 — T4 Madrid Concentration Risk (Single SPOT Provider) The Madrid T4 SPOT market has one provider. A routine capacity reallocation decision by that provider — not a strategic exit, just a quarterly fleet rebalancing — could vaporize the entire SPOT market. Any workload committed to Madrid SPOT capacity should maintain cross-region fallback to an established EU market (Frankfurt, Amsterdam). The ON_DEMAND market with 2 providers is more robust, but still thin by any standard.

R3 — TPU V5 LitePod Step-Down Risk GCP's staircase pricing model for TPU SPOT means the $0.023/hr Virginia and $0.003/hr Singapore floors could step down again in approximately 2–4 weeks. For teams deploying training workloads on this arbitrage, the risk is not preemption but pricing stepping below the cost of checkpoint overhead — at $0.003/hr, cost is already near zero, but GCP could theoretically withdraw the tier entirely if V6E supply is reallocated. Monitor weekly.

Near-Term (30–90 days)

R4 — AWS re:Invent Inferentia3 Announcement: Transition Cliff for Existing Users If Inferentia3 is announced at re:Invent (expected November 2026), the transition playbook from AWS will be to hold Inferentia1/2 on-demand prices stable for 90–180 days while making new capacity available only on Inferentia3 instances. This creates a 90-day window where Inferentia1/2 SPOT capacity may expand temporarily as AWS shifts engineering focus — potentially reversing the current SPOT tightening. Operators who lock RI contracts after an Inferentia3 announcement may find themselves in gen-1 hardware at gen-1 pricing just as workload migration pressure builds. The optimal window is before the announcement.

R5 — Pennsylvania PUC Cost-Allocation Requirement Pennsylvania's Public Utility Commission issued guidance on September 5, 2026 requiring datacenters with peak demand over 25 MW to execute commission-jurisdictional cost-allocation agreements. This formalizes cost-sharing for grid upgrades and will affect every hyperscaler and neocloud operating in PA. The direct pricing implication: operating cost increases in PA will ripple into cloud pricing for US-East (Philadelphia/Pittsburgh) region instances over 2–4 quarters. Monitor for similar state-level cost-allocation moves in Virginia (Loudoun County, where datacenters generate 45% of local tax revenue) — that would be a far larger market impact.

R6 — Nvidia DOJ Antitrust Escalation The DOJ investigation into Nvidia's $17B Groq licensing deal is at an early stage, but the mechanism of concern — exclusive NVLink Fusion access — directly affects the custom silicon ecosystem. If the DOJ forces Nvidia to open NVLink Fusion to non-preferred partners, it could accelerate d-Matrix, Marvell, and Broadcom XPU integration timelines. If the deal is blocked entirely, it fragments the inference silicon stack in ways that temporarily benefit pure-GPU providers. Either resolution has equity implications for NVDA, AVGO, and MRVL. No timeline for resolution is publicly available.

Structural (6–24 months)

R7 — Power Grid Constraint as a Regional Pricing Floor Bloom Energy's Q2 2026 revenue of $1.065B (+166% YoY) and a $20B backlog driven entirely by hyperscaler demand for on-site generation is the most concrete evidence that grid connection timelines now exceed datacenter construction timelines in US and European markets. This structural bottleneck will create durable regional pricing premiums in any geography with constrained grid access — US-East Virginia (already 45% of Loudoun County tax base), Texas, and emerging European markets. The inverse: regions with nuclear baseload (Finland — Google's 22-year Loviisa contract) or rapid on-site generation buildout (Australia — Nvidia's 2GW partnership, SharonAI's planned 68,000 GPU fleet) will become structural pricing discounters over 18–36 months. Geographic exposure to power-constrained regions is a medium-term cost risk for any compute operator.

R8 — China Domestic GPU Supply Catching Up on Gross Margin Trajectory Biren Technology's H1 2026 revenue of $183.9M vs. $8.7M in H1 2025 (+2,000% YoY) at 42.7% gross margins is the most significant competitive data point from China this cycle. Nvidia shipped 2.2 million AI GPUs to China in H1 2025 at $12–15K/unit before the May 2025 export ban; that demand is now being served domestically at accelerating scale. The risk for global GPU pricing is not near-term (Biren and peers are still orders of magnitude below Nvidia's capacity and software ecosystem depth), but the gross margin trajectory (42.7% for a startup with 2,000% growth) suggests Chinese domestic AI silicon is not a charity project — it is building toward competitive unit economics. On a 24-month view, watch for Chinese domestic GPU pricing to influence international SPOT market benchmarks via Singapore and Hong Kong regional hubs.