Nvidia Cuts Authorized Asian Buyers by Half, New York Pauses Data Centers Above 50MW, and $2.55B Flows into Inference Chips in One Day — Compute Supply Constraints Tighten from Politics, Power, and Capital at Once

After last week’s dual releases of GPT-5.6 and Grok 4.5, the model layer was relatively quiet this week. The center of gravity shifted to the compute supply side, and all three major developments pointed to the same issue: supply constraints. Politically, Nvidia cut more than half of its authorized Asian buyers under a new equity-ownership-based review regime, while Huawei announced that the physical Ascend 950 SuperNode system would make its public debut at WAIC this weekend. Export controls and domestic substitution each moved further in the same week. On the power side, New York State signed the first state-level data center moratorium in the United States, pausing approvals for new projects above 50MW for one year. On the capital side, $2.55 billion flowed into inference chip companies in a single day on July 7; one week later, the Bank for International Settlements released a report warning that the debt structure behind AI infrastructure investment may be more dangerous than any previous technology boom. On the regulatory side, the EU has formally completed the legislative process for delaying high-risk obligations, but the August 2 transparency obligations remain on schedule. For companies operating in the EU, the genuinely urgent compliance deadline has not changed.

Major Developments

Nvidia Cuts Asian Buyers by Half and Ascend 950 Makes Its Physical Debut: Controls and Substitution Both Move Further

On July 14, Nvidia began enforcing stricter export compliance reviews for Asian customers, establishing a procurement eligibility whitelist. More than half of its previous customers failed the reassessment and lost direct purchasing eligibility for AI chips. This tightening implements new rules introduced in late May: for advanced AI chip export licensing, the assessment standard has shifted from the delivery location to equity ownership. Entities whose ultimate parent company is registered in mainland China or Macao require a license regardless of whether the receiving entity is located in Singapore, Malaysia, or elsewhere. The backdrop is the March indictment by U.S. prosecutors of three individuals, including a Supermicro co-founder, accused of smuggling approximately $2.5 billion worth of Nvidia chips to China through Southeast Asian intermediary companies. Once transshipment loopholes were closed, the entire Asian channel structure had to be screened again. In parallel, Bloomberg reported that a small number of H200 chips had been shipped to Chinese customers after receiving U.S. authorization. Government officials described the volume as “trivial,” making its symbolic meaning larger than its supply impact.

On the other side of the same week, Huawei announced on July 13 that the new-generation Ascend SuperNode, the Atlas 950 SuperPoD, will make its first public appearance at the World Artificial Intelligence Conference from July 17 to 20. The system supports up to 8,192 Ascend 950DT cards connected through Huawei’s self-developed Lingqu 2.0 all-optical interconnect protocol, with interconnect bandwidth reportedly around 15 times higher than traditional protocols and single-hop communication latency reduced from 2 microseconds to 200 nanoseconds. Each card is equipped with 144GB of HBM, or high-bandwidth memory that vertically stacks multiple memory dies close to the processor package, 4TB/s of memory bandwidth, and new support for low-precision data formats such as FP8 and MXFP4. For multinational enterprises, the two developments should be read together: compute procurement in Asia now needs to reserve time and uncertainty for compliance reviews, and supplier eligibility may change overnight; meanwhile, China’s regional technology stack is accelerating toward an independent Ascend-centered ecosystem. AI architecture choices for China operations will increasingly diverge from the global primary stack and should be planned separately across two hardware ecosystems.

New York Hits Pause: Power Constraints Become State-Level Policy for the First Time

On July 14, New York Governor Kathy Hochul signed an executive order pausing permit approvals for new data centers requiring more than 50 megawatts of power for up to one year. New York thereby became the first U.S. state to impose a state-level data center moratorium. Fifty megawatts is roughly the power scale of a mid-sized AI data center. Projects that have already received permits are exempt. The pause period will be used to assess the environmental impact of data centers and develop new regulatory standards. Supporting measures include a proposed “Grid Acceleration Fund,” which would require data centers to invest in aging in-state grid infrastructure, and a community investment framework to be issued within 60 days, helping local governments negotiate returns such as infrastructure, jobs, and childcare.

The signaling effect of this event is greater than its direct impact — New York was not originally a major hub for hyperscale data centers. The real issue is precedent. As disputes over data centers driving up residential electricity prices continue to spread across multiple U.S. states, New York has put the option of “pause first, set rules later” on the table. Lawmakers in other states now have a template they can cite directly. For enterprises, compute supply in the U.S. East region will not be affected in the short term. But when selecting data center regions and signing long-term compute contracts, state-level policy risk now needs to be written into the assessment. The politicization of power approvals means the geographic distribution of compute supply will be more dispersed than pure economic logic would predict, making cross-region and even cross-border deployment flexibility more valuable.

Nebius Shows Both Sides in One Week: A Neutral Compute Provider Gains a New Customer and Sees a New Competitor

On July 14, open-source model company Reflection AI signed a compute contract worth approximately $1 billion with Nebius, running through 2029. Reflection will gain access to Nvidia’s latest GB300 compute, the current flagship rack-scale platform. The company, founded in 2024, had signed a similar compute agreement with SpaceX only a few weeks earlier. Together with its previous $1 billion compute arrangement with Nebius, these contracts are turning “do not build, rent from multiple sources” into a standard playbook for open-source model companies. For Nebius, this order follows its March agreement with Meta for up to $27 billion in supply over five years, giving its customer roster a weight comparable to first-tier cloud providers.

But the same company had just experienced the other side of the market. After Meta’s self-built cloud business — the Meta Compute plan covered in last week’s brief — was reported, Nebius lost roughly $12 billion in market value in a single day. The market’s concern was straightforward: if the largest customer starts selling compute itself, neutral suppliers may be squeezed from both ends. Taken together, these two sides reflect the real condition of the emerging neocloud sector: model companies on the demand side continue to lock in compute, while the largest buyers on the supply side may become competitors at any time. The procurement lesson for enterprises is clear: the compute supplier landscape is far from settled. For three- to five-year long-term contracts, counterparty risk deserves serious assessment. The cost of splitting contracts across multiple suppliers may be lower than the cost of betting on the wrong one.

$2.55B Enters Inference Chips in One Day, and One Week Later the Central Bank for Central Banks Issues a Warning

On July 7 alone, three inference chip companies secured a combined $2.55 billion: SambaNova completed the first close of a $1 billion Series F at an $11 billion valuation; Positron was reported to be in talks to raise up to $750 million at a valuation of up to $5 billion; and China’s Iluvatar CoreX completed a share sale exceeding $800 million. Together with broader first-half figures — U.S. startups raised $412.7 billion in H1, with 86% flowing to AI companies — capital’s bet on the inference layer is still accelerating. The logic is consistent with OpenAI’s release of its in-house inference ASIC two weeks ago and reports that DeepSeek is developing its own inference chip: inference now accounts for the bulk of AI compute demand, and whoever can reduce inference cost will have a business.

One week later, on July 14, the Bank for International Settlements (BIS), the cooperative organization of central banks, released a research report that poured quantitative cold water on the boom. It found that the scale of AI infrastructure investment could exceed past technology booms that ended in sharp market corrections, and that this cycle is distinctive because debt is more deeply involved. Lending and cross-shareholding relationships between hyperscale cloud providers and model companies mean that if productivity gains fail to keep up with investment, the impact may not stop at stock prices and could transmit into broader financial instability. The two developments do not contradict each other. They are two readings of the same phenomenon: money is still entering the market, and more of that money is borrowed. For enterprises, this does not change any short-term procurement decision, but the inference is worth remembering: if debt-driven consolidation later emerges on the supply side, contracts tied to a single compute or model supplier may become liabilities at the worst possible moment.

EU High-Risk Delay Is Settled, but the August 2 Transparency Obligation Has Not Been Delayed

The EU AI Act simplification package, the Digital Omnibus on AI, has completed the final legislative steps: the European Parliament voted to approve it on June 16, the Council gave final approval on June 29, and the text is expected to be published in the Official Journal in July and enter into force three days later. The core change is a substantial postponement of compliance obligations for high-risk AI systems. Annex III scenarios — including recruitment, credit scoring, law enforcement, education, and border control — move from August 2, 2026 to December 2, 2027. Annex I systems embedded in regulated products are pushed further to August 2, 2028.

The point easily obscured by this delay is what has not been delayed: Article 50 transparency obligations still take effect on August 2, 2026 according to the original timetable. Users must be informed when they interact with AI systems, and AI-generated or manipulated content must be identifiable. This is 18 days away, and it is a generally applicable obligation for all companies offering AI functions in the EU, not limited to high-risk scenarios. Law firm alerts have also highlighted a tail risk: the delay has no legal effect until the text is formally published in the Official Journal. If the process slips past August 2, the original timetable would apply as written. For enterprises, the remaining work in July is concrete: inventory AI interaction points and content generation pipelines facing EU users, confirm that notification and labeling mechanisms are in place, and retain evidence-ready records. For companies running AI traffic through AgentsFlare, request-level audit trails and call records by user and endpoint can directly serve as the underlying data for this type of compliance evidence. Along the same line of compliance productization, Anthropic launched self-serve HIPAA configuration for enterprise and API customers on July 15, allowing administrators to review the BAA online and enable it with one click. Also note that Claude Opus 4.1 was announced as deprecated on June 5 and will officially retire from the API on August 5; workloads still using it should migrate as soon as possible.

Compute Upstream: Spreads in the GPU Rental Market Matter More Than Prices Themselves

July GPU rental market analyses, including those from Thunder Compute and other tracking sources, show that the most important signal is the spread structure. For the same H100, on-demand rental prices across suppliers range from $1.49 to $6.98 per hour. Dedicated GPU clouds are generally 50–75% cheaper than hyperscale cloud providers. Prices for mainstream cards such as the H100, H200, and A100 have remained broadly stable throughout the year. B200 on-demand prices are down around 7% year over year, while the newest-generation cards such as B300 and MI300X have nearly doubled within a year. The market transmission path is clear: the current elasticity in inference cost comes mainly from procurement channels rather than the hardware itself. Switching suppliers can reduce cost far more than waiting for the next-generation card. Together with the supplier expansion implied by this week’s Meta Compute and Nebius contract wave, the buyer’s negotiation environment in the rental market is improving — provided that enterprise workloads have the engineering capability to migrate across suppliers.

Overall Assessment: A Week When Constraints Became Visible

First, all three categories of compute supply constraints became visible in the same week. Control constraints, represented by Nvidia’s whitelist; power constraints, represented by New York’s moratorium; and capital constraints, represented by the BIS debt warning, all existed before, but mainly as background risks. This week, each became a specific fact with a document, a list, or a data point. The common implication for enterprises is that compute supply planning has moved from “procure on demand” toward a longer cycle that requires anticipating policy direction and supplier financial health.

Second, the divergence between U.S. and China hardware stacks is accelerating in lockstep. It is not a coincidence that Nvidia’s Asian buyer list was cut in half in the same week that the physical Ascend 950 system was set for its first public appearance. Every tightening of controls raises the certainty premium of the domestic stack. For multinational enterprises with substantive operations in China, parallel planning across two hardware ecosystems and two model ecosystems is no longer a contingency plan. It is already underway.

Third, the correct reading of the compliance timetable is layered. The EU has delayed the heaviest high-risk obligations by 16 months, which is real pressure relief. But the August 2 transparency obligation and the August 5 retirement of Opus 4.1 point to another reality: delays usually occur for the heaviest obligations, while lighter and more certain deadlines arrive on time. Compliance resources should be prioritized accordingly.

References

Yahoo Finance — Nvidia Tightens Asian Customer Approvals as US Chip Export Controls Intensify: https://finance.yahoo.com/technology/ai/articles/nvidia-tightens-asian-customer-approvals-105148051.html — 2026-07-14

TechTimes — Nvidia Cuts Over Half of Asian AI Chip Buyers as BIS Compliance Net Widens: https://www.techtimes.com/articles/320527/20260715/nvidia-cuts-over-half-asian-ai-chip-buyers-bis-compliance-net-widens.htm — 2026-07-15

Bloomberg — Small Amount of Nvidia AI Chips Sold to China Via US License: https://www.bloomberg.com/news/articles/2026-07-14/small-amount-of-nvidia-ai-chips-shipped-to-china-with-us-license — 2026-07-14

Sina Finance — Ascend 950 SuperNode Physical Debut Approaches, Domestic AI Compute Clusters Enter the 10,000-Card Era: https://finance.sina.com.cn/stock/t/2026-07-14/doc-inihtrnk8615689.shtml — 2026-07-14

Axios — N.Y. Gov. Kathy Hochul signs data center moratorium executive order: https://www.axios.com/2026/07/14/ny-gov-kathy-hochul-data-center-moratorium-executive-order — 2026-07-14

Washington Post — New York becomes first state to impose data center moratorium: https://www.washingtonpost.com/technology/2026/07/14/new-york-becomes-first-state-impose-data-center-moratorium/ — 2026-07-14

Fortune — NYS Gov. Hochul's data center moratorium includes a new model for funding AI infrastructure: https://fortune.com/2026/07/14/new-york-governor-kathy-hochul-data-center-moratorium-ai-infrastructure/ — 2026-07-14

TechCrunch — Reflection inks $1B compute deal with Nebius: https://techcrunch.com/2026/07/14/reflection-inks-1b-compute-deal-with-nebius/ — 2026-07-14

Nebius — Nebius signs new AI infrastructure agreement with Meta (March, background): https://nebius.com/newsroom/nebius-signs-new-ai-infrastructure-agreement-with-meta — 2026-03-16

TECHi — Nebius Lost $12 Billion in a Day. Meta's Cloud Plan Explains Why: https://www.techi.com/nebius-stock-meta-cloud-plan/ — 2026-07

StartupHub.ai — AI chip inference wars: three companies raise $2.55B in one day: https://www.startuphub.ai/ai-news/ai-news/2026/ai-chip-inference-wars-july-8-2026 — 2026-07-08

Bloomberg — AI Infrastructure Spending May Outpace Past Tech Booms, BIS Study Finds: https://www.bloomberg.com/news/articles/2026-07-14/ai-investment-race-could-turn-debt-fueled-boom-to-bust-bis-says — 2026-07-14

INSIGHT EU MONITORING — EU Council gives final approval to AI Act simplification under Omnibus VII: https://ieu-monitoring.com/editorial/eu-council-gives-final-approval-to-ai-act-simplification-under-omnibus-vii/1244434 — 2026-06-29

Latham & Watkins — AI Act Update: EU Resolves to Change Rules and Extend Deadlines: https://www.lw.com/en/insights/ai-act-update-eu-resolves-to-change-rules-and-extend-deadlines — 2026-07

Holland & Knight — U.S. Companies Face EU AI Act's Possible August 2026 Compliance Deadline: https://www.hklaw.com/en/insights/publications/2026/04/us-companies-face-eu-ai-acts-possible-august-2026-compliance-deadline — 2026-04

Claude Help Center — Release notes (self-serve HIPAA configuration, 7/15): https://support.claude.com/en/articles/12138966-release-notes — 2026-07-15

Claude Platform Docs — Model deprecations (Opus 4.1 deprecated 6/5, retiring 8/5): https://platform.claude.com/docs/en/release-notes/overview — 2026-06-05

Thunder Compute — AI GPU Rental Market Trends (July 2026): https://www.thundercompute.com/blog/ai-gpu-rental-market-trends — 2026-07

IntuitionLabs — H100 Rental Prices Compared: $1.49–$6.98/hr Across 15+ Cloud Providers: https://intuitionlabs.ai/articles/h100-rental-prices-cloud-comparison — 2026-07

GuruFocus — NVIDIA Tightens Compliance for AI Chip Sales Amid U.S. Export Controls: https://www.gurufocus.com/news/8957021/nvidia-tightens-compliance-for-ai-chip-sales-amid-us-export-controls-nvda — 2026-07-14