Daily Investment Brief — September 20, 2026

Signwl Daily Compute Brief — 20 September 2026

News Interrogation Mode | Themes Confirmed, Contradicted, or Silent

Market Pulse

The compute market has fractured into two distinct economies that share a hardware SKU but little else. At the frontier — H100 and B200 in capacity-constrained geographies — pricing is firm to rising, exemplified by Nebius's pre-announced ~20% H100/H200/B200 hike effective October 1 and São Paulo's H100 SPOT holding at $9.35/GPU-hr with a paper-thin 16% discount in a single-provider market. Below that layer, a broad mid-tier deflation is wiping 40–70% off A100 40GB, A10G, L4, and V100 SPOT pricing across US and European markets over 30 days — not because workloads evaporated, but because the inference computation itself is rotating: purpose-built ASIC infrastructure (Inferentia SPOT, up +112–140% over 30d in Virginia and Oregon after two consecutive troughs at $0.034/chip-hr) is reclaiming workloads that temporarily migrated to ultra-cheap GPU spot. This is a sub-market decoupling story with a geographic dimension attached: the "GPU scarcity" narrative that dominates the press is simultaneously true (Tokyo H100 SPOT, 2 providers, $2.79/hr, narrowing spread) and false (Frankfurt H100 SPOT, $0.61/hr, 372% spread between min and max, a liquidation market masquerading as a scarcity one). The dominant pricing regime today is generational bifurcation compounded by geographic fragmentation — and the consensus narrative has conflated these into a single "AI demand is insatiable" thesis that the data repeatedly does not support in aggregate.

Key Movers

Discard B200 Stockholm entirely. The overnight +88.5% move is the pipeline's one flagged anomaly (0.04% anomaly rate across 2,734 tickers on Sep 20). The history shows a step-collapse from $4.48/hr in April to $1.50–$1.71/hr on near-zero observations before the single-SKU re-entry. This is not a market price.

Investable Insights

H1 — H100 Spot Is Seventeen Separate Markets, and the Arbitrage Is Widening Into a Structural Feature Confidence: 5 / 5

Thesis: The dominant press narrative — "GPU scarcity is squeezing AI builders" — is simultaneously true and false depending entirely on which of 17+ observable H100 SPOT markets you examine. The bifurcation has a 58x price range across identical hardware: Amsterdam at $0.16/hr versus São Paulo at $9.35/hr. This is not a temporary inefficiency that arbitrage will close. The $9.35 São Paulo price reflects a market with a single provider, zero competitive tension, and a thin observation pool (n=2), while the $0.61 Frankfurt price exists precisely because multi-provider depth (n=3+ with a 372% min-max spread) has produced a liquidation dynamic. Neither price is "correct" — they are each locally rational given the supply depth behind them.

The structural persistence mechanism is threefold. First, physical supply takes 3–5 years to build — the entire Northeast Asian gigawatt pipeline (South Korea 8.4 GW Phase 1, Japan's Goodman Tsukuba 1 GW Phase 1) has no meaningful delivery before 2028, leaving Tokyo H100 SPOT at $2.79/hr with 2 providers and a narrowing 10.3% spread as the tightest visible multi-provider market in the dataset. Second, data sovereignty, latency requirements, and hyperscaler contractual commitments prevent instant workload repatriation to Frankfurt's $0.61/hr pricing. Third, the Western deflation cycle (A100 40GB Montreal -49% over 30d, A10G Stockholm -42%) is depressing the floor of the supply curve faster than APAC markets can add competitive providers. The result is not market inefficiency but market segmentation hardening into permanent price tiers.

Critically, the KL ($1.28/hr) and Singapore ($1.03/hr) single-provider tickers reveal a nuance the original hypothesis understated: these thin APAC markets are cheap relative to US pricing, not expensive — aggressive neocloud entrants buying market share. The APAC premium story is specifically a Japan/Korea story, not pan-APAC. The Nebius October 1 price hike (+20% on H100/H200/B200) — which triggered a 9% NBIS equity move on September 17 — independently corroborates that frontier pricing in at least one European neocloud market is not in deflation.

Key Evidence:

1. São Paulo H100 SPOT at $9.35/hr (16% discount, n=1 provider, n_obs=2); opened Jan at $5.83/hr, peaked April at $9.80/hr on FX/supply shock — 90d price history confirmed (unified model ticker, Sep 20)
2. Amsterdam H100 SPOT at $0.16/hr; Frankfurt at $0.61/hr with 372% min-max spread across 3 providers — same hardware, 58x price differential (cross-provider tickers, Sep 20)
3. H100 SPOT Tokyo: $2.79/hr, 10.3% spread, 2 providers, spread narrowing -2.5% over 30d — a thin but stabilizing market, not a glut (ticker data, Sep 20)
4. South Korea 8.4 GW datacenter buildout: zero groundbreaking before H1 2028; KEPCO transmission network incomplete until 2036 for full grid (datacenter tracker, Sep 20)
5. Japan active construction: SoftBank Tomakomai 300 MW (2026), SoftBank Sakai 140 MW (2026), Goodman Tsukuba Phase 1 50 MW (2028), NTT Data Osaka North (2027) — under 500 MW operational before 2028 (datacenter tracker, Sep 20)
6. Nebius ~20% H100/H200/B200 price hike announced effective October 1, NBIS +9% on announcement; 25% AMD EPYC CPU hike; 41% memory hike — frontier operators raising prices, not cutting (GuruFocus, Sep 17)
7. NatEx fields null for all APAC H100 SPOT tickers — structurally expected for single-provider thin markets lacking matched-pair catalog variety; confirms these are not model artifacts but genuine supply-structure reads (unified model diagnostics, Sep 20)

Implied Action: For AI workload operators: latency-tolerant training workloads should be migrated from APAC/Latam to US Central (H100 SPOT Iowa at $1.54/hr) or Frankfurt ($0.61/hr) — the 87% cost differential versus Dubai ($2.87/hr) or the 93% differential versus São Paulo represents immediate, actionable savings with no hardware or provider change needed. For infrastructure investors: APAC-focused compute operators with existing power-secured capacity — specifically SoftBank (9984.T) ahead of its October GB200 NVL72 commercial launch — have 2–4 years of structural pricing power that Western-centric neocloud operators cannot compete against. Avoid US/EU neocloud operators whose H100 inventory is exposed to the Frankfurt/Amsterdam liquidation dynamic. Monitor the São Paulo H100 SPOT ticker (H100|SPOT|br-saopaulo) for any second-provider entry — that event alone would compress the price 25–35% and is the clearest signal that the Latam premium is ending.

H2 — CoreWeave Is a Debt-Financed Long-Duration Bet Against a Spot Market That Is Already Below Its Reservation Price Confidence: 4 / 5

Thesis: CoreWeave's financial architecture is a leveraged carry trade: borrow cheaply (convertible notes at 2.875%, long-dated to 2033), deploy capital into GPU infrastructure at reserved rates ($40M/MW/yr per management's own reference), and earn the spread versus spot. The problem is that spot has moved adversely and aggressively. H100 SPOT in US East is at $1.54–$1.92/GPU-hr. At 8 GPUs per standard H100 server and ~50% average utilization for a spot pool, the implied revenue per MW from spot is roughly $3.5–$5M/MW/yr — a 6–8x shortfall versus the $40M/MW/yr reserved-rate target. That gap is the mechanism by which spot deflation threatens CoreWeave: not through immediate contract breach, but through renewal pricing compression as customers who signed 3–6 month contracts come back to the table against a spot benchmark that has fallen 60–80% since many of those contracts were written.

CRWV's Q2 2026 financials confirm the burn trajectory: FCF was -$5.7B in Q2 alone (capex $6.4B, operating cash flow $679M), with TTM FCF at -$9.1B. The acceleration is the alarming feature — Q1 was -$4.7B, Q2 was -$5.7B, implying the capex program is still ramping, not decelerating. Against $51.6B in total debt and $5.5B in cash (current ratio 0.455 — technically current-liability-insolvent), the just-closed $3.7B convertible note offering (upsized from $3.0B, September 17, conversion at $97.85 against a stock at $79.88) is dilutive equity at a 22.5% conversion premium that the market has not yet embraced. The stock's -30.7% three-month move against a $144 analyst consensus reflects partial — but not complete — discounting of this exposure.

The bear case requires precision. The "$25B+ in additional Q3 customer commitments" disclosed in late-August news is real and meaningful — it is not consistent with a business already in churn crisis. CoreWeave has also expanded contracted power capacity to 4.2 GW by August, up 13.5% in one quarter. The bear thesis is therefore a maturity cliff, not an immediate default: when those Q3 contracts roll at renewal (3–6 months → Q4 2026/Q1 2027), what is the renewal rate? At $40M/MW/yr, the model works. At $20M/MW/yr — the rate implied by current spot pricing at full-utilization economics — revenue per MW halves at renewal. The November 11 Q3 earnings are the catalyst that will reveal whether the backlog maturity structure is compressing.

Key Evidence:

1. CRWV Q2 2026 FCF: -$5.7B (capex $6.4B vs. operating CF $679M); Q1 2026 FCF: -$4.7B; TTM FCF: -$9.1B — burn rate accelerating (CRWV 10-Q/fundamentals data, Sep 20)
2. CRWV total debt $51.6B, cash $5.5B, current ratio 0.455; D/E ratio 1,027x (fundamentals data, Sep 20)
3. $3.7B convertible note closed Sep 17 (upsized from $3.0B), 2.875% coupon, conversion at $97.85 vs. stock at $79.88 — 22.5% premium, dilutive (CRWV 8-K, Sep 17)
4. H100 SPOT US East: $1.54–$1.92/GPU-hr vs. CRWV management's implied reserved rate of $40M/MW/yr ($9–12/GPU-hr equivalent at full utilization) — a 4–6x spot-vs-reserved gap that compounds at each renewal cycle (unified model ticker, Sep 20)
5. A10G SPOT Stockholm -42%, A100 40GB SPOT Montreal -49% over 30d — mid-tier GPU floor eroding below CRWV's older fleet segment (top movers, Sep 20)
6. "$25B+ in additional Q3 customer commitments," power capacity 3.7 → 4.2 GW in one quarter — modifies but does not invalidate the bear; contract maturity structure is the unknown (news feed, Aug/Sep 2026)
7. CRWV -30.7% over 3 months at $81 vs. $144 analyst consensus, $39 analyst low; APLD -42%, CIFR -36%, BTDR -37% — neocloud sector repricing underway (equities feed, Sep 20)
8. NSCALE S-1 filed September 18 — new entrant adding supply into an already-deflating spot market (SEC feed, Sep 18)
9. Purchase commitment footnotes for CRWV returned empty from the pipeline — the single most critical data gap for precise bear sizing (purchase_commitments query, Sep 20)

Implied Action: Position for the November 11 earnings reveal. The most asymmetric short is sized around three specific data reveals in the Q3 10-Q: (a) total purchase commitment outstanding dollar value and maturity buckets — if backlog maturities are clustered in Q4 2026/Q1 2027, the renewal-price risk is immediate; (b) gross margin trend — any compression below 65% (currently 67.4%) signals pricing pressure has already arrived; (c) debt covenant headroom disclosures in the notes. The $39 analyst bear target requires a covenant-stress scenario; a "soft-landing" bear case at $50–55 is more realistic if $25B Q3 backlog converts at near-original rates. The $144 bull case requires renewal pricing to hold despite a spot market that has moved 60–80% against it — a heroic assumption. Do not treat this as binary; enter with a tranche ahead of November 11 and size the remainder on the Q3 reveal. Avoid being short ahead of any H100 on-demand price increase by AWS/Azure/GCP — that event would partially validate the bull case by lifting the market floor.

H3 — AWS Is Recapturing Inference Workloads on ASIC Infrastructure, and the Cheap-GPU-Inference Window Is Closing Confidence: 4 / 5

Thesis: The most analytically revealing dataset in this briefing cycle is not a headline number — it is the four-phase price history of the AWS inf1.24xlarge SPOT ticket in Oregon from January 8 to September 20, 2026. From $0.162/chip-hr in January, Inferentia SPOT decayed in a nearly linear 79% collapse to a trough of $0.034/chip-hr by April 6 as inference workloads migrated to ultra-cheap GPU spot (L4, A10G, A100 40GB were simultaneously in deep deflation). Then came a double-bottom structure — a false recovery to $0.074 in April, a second collapse to the identical $0.034 trough by August 7, and finally, a sustained 196% recovery over 43 days to today's $0.105/chip-hr with no catalog churn (n_instance_types=8 stable throughout; all 50 observed instances are inf1.24xlarge, no SKU turnover). The double-trough is diagnostically important: the April recovery failed because GPU spot was still cheap enough to compete. The August recovery is sustaining because GPU spot has stabilized and Inferentia's price/performance advantage for latency-sensitive inference is becoming decisive again.

This is not deliberate AWS capacity throttling — AWS cannot set spot prices in real-time, and the stable catalog composition (no pool-size manipulation visible) rules out supply-side engineering. It is organic inference demand rotating back onto purpose-built silicon as the GPU spot window closes. The Baseten engineering data on diffusion speculative decoding (3x+ improvement on a single B200 running Qwen 3 8B) and KV cache compression confirms that inference optimization is accelerating in ways that favor high-utilization reserved ASIC capacity over burst GPU spot. Inferentia in Tokyo, at $0.369/chip-hr — 3.5x the Oregon price — adds an APAC dimension to the thesis that is fully consistent with the broader geographic scarcity pattern in H1.

The practical implication is direct: AI API providers who built inference pipelines on the assumption of near-zero GPU spot — exemplified by DeepSeek's $0.15/M input token pricing on its 552B MoE V4.1-Flash — will face COGS pressure as both the ASIC and GPU inference floors rise simultaneously. The October 1 Nebius hike on H100/H200/B200 ($0.069 SPOT → roughly $0.083/hr post-hike) compounds this dynamic on the GPU side. The ultra-cheap inference arbitrage window that opened in spring 2026 is closing from two directions at once.

Key Evidence:

1. Inferentia SPOT Oregon: +140% over 30d, $0.035 → $0.105/chip-hr since Aug 8 — confirmed 196% recovery over 43 days, zero catalog churn, n_instance_types=8 stable throughout (underlying instances inf1.24xlarge, Sep 20)
2. Inferentia SPOT Virginia: +112% over 30d, $0.041 → $0.087/chip-hr — independent corroboration in separate region (ticker data, Sep 20)
3. Inferentia SPOT Tokyo: $0.369/chip-hr, +41% over 30d — APAC inference premium at 3.5x Oregon pricing (ticker data, Sep 20)
4. Double-trough structure: Trough 1 at $0.034 (Apr 6), false recovery to $0.074 (Apr 18), Trough 2 at $0.034 (Aug 7), sustained recovery since — mechanistically consistent with GPU spot competition failing to hold inference workloads through July (90d ticker history, Oregon, Sep 20)
5. inf1.24xlarge launched December 2019; still active in AWS spot catalog 6.8 years later; inf2 instances is_active: false in spot catalog — deliberate product-tier preservation of v1 ASIC for spot inference workloads (component tracker / depreciation data, Sep 20)
6. Baseten inference engineering: diffusion speculative decoding delivering 3x+ on single B200 for Qwen 3 8B; KV cache compression moving to learned representations — inference optimization accelerating toward high-utilization ASIC patterns (finance.biggo.com, Sep 19)
7. Anomaly diagnostics: 0.04% anomaly rate (1 anomaly, B200 Stockholm); Inferentia ticker is clean — not a model artifact (BQ anomaly_diagnostics, Sep 20)
8. No UPCOMING_PRICE_CHANGE alert on Inferentia on-demand — the recovery is a spot market signal, list-price confirmation still pending (price_alerts query, Sep 20)

Implied Action: Watch October 1 as a dual-signal checkpoint: (1) whether Nebius's H100/H200/B200 price hikes hold (customer acceptance = frontier pricing power confirmed); (2) whether Inferentia SPOT continues its recovery trajectory past $0.110/chip-hr (which would put the spot discount below 72% for the first time since Q1). If both materialize simultaneously, the cheap-inference window has definitively closed and AI API providers running GPU-spot inference need to reprice or re-platform within 1–2 quarters. AWS infrastructure margin is a direct beneficiary. For inference-first AI companies pricing below $0.20/M input tokens on GPU spot, this is a model risk, not just a cost risk — the pricing itself becomes structurally unviable if COGS move 30–50% against them. The Inferentia on-demand list price is the cleanest leading indicator to watch. Any AWS-published price increase on inf1 on-demand would confirm pricing power has definitively returned to the ASIC inference tier.

H4 — Northeast Asia GPU Supply Is Power-Gated Until at Least 2028, Sustaining a 45–65% Regional Premium Confidence: 4 / 5

Thesis: The $2.79/hr H100 SPOT price in Tokyo — 45% above US East's $1.92/hr for identical hardware — is not a temporary anomaly awaiting arbitrage. It is the market price of a regional supply-demand imbalance that will not structurally resolve before 2028 at the earliest, and in South Korea, not before the mid-2030s for large-format capacity. The supply pipeline is almost entirely paper: South Korea's entire 8.4 GW Phase 1 buildout has no groundbreaking before H1 2028, and the Yongin 15 GW "megacluster" is a semiconductor fab, not a GPU cloud — it adds zero GPU-hours of compute supply. Japan is somewhat better positioned: SoftBank's Tomakomai (300 MW) and Sakai (140 MW) facilities are operational-target 2026, and both are power-secured on existing infrastructure — Tomakomai on Hokkaido hydro/solar/wind, Sakai on the repurposed Sharp factory grid connection. But these are small-format deployments absorbed largely by SoftBank's own domestic AI service rollout.

The pricing power dynamic is acute. In Japan, Azure prices H100 at $10.55/hr on-demand versus AWS at $7.53/hr — a 40% Azure premium in the same city that is sustainable only in a thin-supply market where a second provider can price above the first without losing customers to competitive alternatives. This Azure/AWS spread in Tokyo is wider than in any US market. The SoftBank GB200 NVL72 commercial launch in October 2026 is a genuine catalyst: first GB200 availability in Northeast Asia at sovereign scale, into a market where current supply is 2 providers (H100) with a narrowing 10.3% spread. If SoftBank prices GB200 at even $6–8/hr in Tokyo (versus the $4–5/hr range visible in early US markets), it would represent 30–60% geographic pricing power above equivalent US hardware — and would immediately re-rate Japanese cloud infrastructure equity as a premium asset class.

Key Evidence:

1. H100 SPOT Tokyo: $2.79/hr, 10.3% spread (2 providers), spread narrowing -2.5% over 30d — tightest stabilizing multi-provider market in APAC (ticker data, Sep 20)
2. Azure H100 OD Tokyo $10.55/hr vs. AWS $7.53/hr — 40% hyperscaler spread within same city, sustainable only in supply-constrained market (underlying instances, Sep 20)
3. South Korea 8.4 GW national AI data center plan: zero groundbreaking before H1 2028; KEPCO transmission network won't complete until 2036; ~6 GW lacking finalized supply plans (datacenter tracker, Sep 20)
4. Japan operational-target construction: SoftBank Tomakomai 300 MW (2026, 100% renewable, Hokkaido hydro/solar/wind) + Sakai 140 MW (2026, Sharp factory grid) — power-secured precisely by avoiding new grid applications (datacenter tracker, Sep 20)
5. Goodman Tsukuba Phase 1: 50 MW under construction, operational 2028 — only institutionally credible large-format Japan supply addition visible before 2030 (datacenter tracker, Sep 20)
6. Keppel Ansan (Korea): 60 MW, 2030 target — the sole institutionally credible near-term Korea addition at modest scale, power-approved (datacenter tracker, Sep 20)
7. SoftBank GB200 NVL72 commercial service launching October 2026 — first GB200 in Northeast Asia, zero competitive alternative in our supply tracker; strong implied pricing power (news feed / datacenter tracker, Sep 20)
8. Naver GAK Sejong (1 GW) + SK/Nvidia Ulsan (1 GW) + GS Donghae/Dangjin (2.4 GW): all H1 2028 groundbreaking — the institutional demand anchor (Naver DSX Nvidia partnership, Nvidia $1B NAVER equity stake, Brookfield $9B facility) is confirmed but supply timeline is not accelerating (datacenter tracker / news feed, Sep 20)

Implied Action: SoftBank (9984.T) is the highest-conviction near-term APAC compute equity catalyst. The October 2026 GB200 NVL72 launch gives SoftBank a 12–18 month monopoly window on the newest frontier hardware in the tightest large GPU market in the world. Western analysts covering SoftBank typically focus on Vision Fund and ARM Holdings exposure — the sovereign AI compute story is under-covered in English-language research. A long position here is a play on pricing power that the pricing data directly corroborates. Monitor the Tokyo H100 SPOT spread monthly — if it widens above 20% (from today's 10.3%), a third provider has entered and the scarcity premium is beginning to erode. If Goodman Tsukuba Phase 1 (50 MW) is announced delayed beyond 2028, the supply gap widens further. CyrusOne-Kansai Electric JV groundbreaking (900 MW, autumn 2026) is the key watch event for Japan supply — if it slips, the 2028 supply thesis tightens.

Inventory & Channel Health

The inventory health composite regime returned no live DIO signals in the current pipeline snapshot — the distributor, buyer, memory, and chip-supplier layers are all absent from this cycle's data. This is a data gap rather than a clean read, and it is the largest single analytical hole in the briefing. However, the trade flow and macro data fill part of the void.

The HS 852351 (flash/SSD) import signal is at extreme levels: +499% year-over-year and +75% over the trailing 3 months as of July 2026, with US flash imports at $7.1 billion in a single month. This is the strongest flash import reading in the dataset, and historically this series leads hardware channel glut formation by 1–2 quarters. The channel dynamic is straightforward: flash goes into SSDs, SSDs go into servers, servers go into data centers — and the current datacenter construction pipeline is absorbing storage ahead of compute cluster completions. The risk materializes if absorption slows. Project Jupiter/Stargate New Mexico — 4 GW planned, $165B committed, named tenants (Oracle/OpenAI), currently 26% complete — is legally blocked on water usage (New Mexico Supreme Court pause) with a gas pipeline delay to February 2027. If that pause extends, the storage already imported in anticipation of that cluster's completion has nowhere to go for quarters longer than modeled.

The HS 8542 (all ICs) signal is flat with a -13% month-over-month print in July — a crucial non-confirmation. If a true AI demand supercycle were pulling silicon broadly, IC imports would be surging alongside flash. The divergence suggests the flash surge is storage-specific datacenter pre-positioning, not a broad demand pull-through. The FRED business inventories-to-sales ratio is 1.3x and flat (3-month momentum -3%), which provides no glut signal yet — but this is a lagging indicator that would not yet reflect July's flash import surge in the channel. The inventory read is: watch the August and September flash import data. If the surge persists into Q3 while Stargate NM remains blocked, the probability of a distributor-channel glut signal emerging in October is elevated.

Supply & Demand Pipeline

Supply — The Three Facilities That Actually Matter in the Next 12 Months:

The Barn/Stargate Michigan (1.4 GW planned, $56B total, Oracle/OpenAI/Related Companies) is the single highest-confidence large-format facility in the US construction pipeline — Building 1 is near complete, with a 19-year DTE Energy power contract. This is the only project of scale where power is confirmed, tenant is named, and construction is visibly progressing. Its full-utilization date matters directly for H100/B200 SPOT pricing in US Central — when this facility absorbs a large GPU cluster, it will pull SPOT pricing upward in the Midwest market (currently H100 SPOT Iowa at $1.54/hr, the global SPOT floor) as available spot pool capacity narrows.

SoftBank's Tomakomai (300 MW, Hokkaido) and Sakai (140 MW, Osaka) are both targeting 2026 completion and are power-secured — the most executable additions in Northeast Asia. Combined they are under 500 MW, which will be absorbed almost entirely by SoftBank's domestic GB200 NVL72 service launch in October 2026. They will not materially improve H100 SPOT availability in Tokyo for third-party buyers.

Stargate NM (4 GW, $165B, 26% complete) remains legally blocked. The New Mexico Supreme Court water-usage pause and the February 2027 gas pipeline delay mean the demand this facility was intended to absorb — the primary named hyperscale training cluster in the US pipeline — is parked. This directly supports the CRWV bear case: Oracle/OpenAI demand that might otherwise have flowed through to reserved capacity contracts is delayed.

Demand — Models Driving the Inference Rotation:

The September model releases confirm the inference rotation thesis in real terms. OpenAI GPT-Live-1 (full-duplex voice, $0.05/minute effective September 10) and Google Gemini 3.8 Live (speech-to-speech, 97 languages, September 15) represent the emergence of real-time multimodal inference as a production workload — a category that demands low-latency, high-utilization inference silicon rather than burst GPU spot. DeepSeek V4.1-Flash ($0.15/M input tokens, 552B MoE with 8–16B active parameters during inference, broadly available from September 10) is the demand-side stress test for the ASIC inference recovery hypothesis: its pricing is only sustainable if GPU spot inference COGS stay near zero, which the Inferentia recovery data suggests will not persist. DeepSeek V4 Flash Vision Exp (763B total parameters, 8–16B active) across 13 distribution channels at commodity pricing reinforces the competitive pricing floor on the API side, even as the cost side is rising.

Pre-Announced Price Changes

No UPCOMING_PRICE_CHANGE alerts exist in the current pipeline window across AWS, GCP, Azure, or Oracle compute or storage. No provider has published a forward effective-date repricing for any instance class.