AI & Compute Infrastructure — September 12, 2026

🇺🇸OpenAI ruled out a 2026 IPO, postponing its listing while prioritizing additional AI-safety work. #

OpenAI CEO Sam Altman said a 2026 initial public offering (IPO) would not be appropriate because the company still had substantial safety-related work to complete. The supplied reports describe this as a decision to defer listing, rather than abandonment of the IPO process.

One report states that OpenAI confidentially submitted IPO documents in June. Although the reports reference 2027 as the next listing timeframe, they also note that timing remained unsettled; no underlying filing, offering terms, valuation, or confirmed listing date is provided.

Altman framed the decision around concerns about the risks associated with increasingly capable AI systems. His reference to even a 10% risk of catastrophic harm being unacceptable was a statement about risk tolerance, not a quantified assessment establishing that such a probability exists.

• OpenAI

🇨🇳Bank of China (中国银行) and partners plan at least RMB300 billion in compute-sector financial support during the 15th Five-Year Plan (十五五). #

Bank of China (中国银行) launched the Yiwang Sanrong compute-finance initiative (一网三融) at the September 12 opening of the 2026 China Computing Conference (中国算力大会). The announcement describes guidance and support from the Ministry of Industry and Information Technology (工业和信息化部) and the People’s Bank of China (中国人民银行).

Participating partners include China Telecom (中国电信), China Mobile (中国移动), China Unicom (中国联通), the People’s Insurance Company of China (中国人民保险集团), China Orient Asset Management (中国东方资产管理股份有限公司), and CDB Leasing (国银金融租赁股份有限公司). The proposed ecosystem brings together banking, insurance, industrial investment, asset management, and financial leasing.

During the 15th Five-Year Plan (十五五), the partners intend to provide at least RMB300 billion in comprehensive financial support to companies across the compute supply chain. Other stated targets include supporting at least 100 key AI computing-center projects, serving more than 3,000 compute-using enterprises, and providing coordinated cross-regional financial services for national compute-interconnection regional nodes.

These figures are program targets, not disclosed completed financing or funds already disbursed. The message does not provide individual loan commitments, interest rates, collateral requirements, borrower allocations, or drawdown conditions.

🇨🇳Space-compute developer Xingce Weilai (星测未来) completed A+ and A++ rounds totaling hundreds of millions of renminbi. #

Xingce Weilai (星测未来) announced completion of its A+ and A++ financing rounds at the space-computing innovation forum of the 2026 China Computing Conference (中国算力大会). The reported hundreds of millions of renminbi is the combined amount across both rounds, not an amount disclosed for each round separately.

The participating investors named in the report are Beichuangtou (北创投), Meridian Capital (华映资本), Xinshang Capital (新尚资本), Xinding Capital (新鼎资本), Jiangsu Kanghua Group (江苏康华集团), Yunsheng Capital (匀升资本), Chenfeng Jishi (晨峰基石), Hainan Jianyue (海南见月), Junyi Capital (君一资本), and Huifengda (惠丰达). Existing shareholders Xichuangtou (锡创投) and Houtian Capital (厚天资本) also made follow-on investments.

The report describes the company’s positioning around full-stack space-compute token operating services. It does not disclose valuation, ownership dilution, financing conditions, deployment dates, orbital hardware quantities, or contracted compute capacity, so the financing announcement should not be treated as evidence of those operating milestones.

• Xingce Weilai (星测未来)

🇨🇳Embodied-model developer Xieyue Intelligence (斜跃智能) closed angel-plus financing worth hundreds of millions of renminbi to fund training and compute infrastructure. #

Xieyue Intelligence (斜跃智能) announced completion of an angel-plus financing round worth hundreds of millions of renminbi. Investors included Linear Capital (线性资本), Junshan Capital (钧山资本), Hongyi Capital (弘颐资本), and Hidden Hill Capital (隐山资本); the reports do not disclose an exact amount or valuation.

Founded in February 2026, the company develops embodied foundation models, with households identified as its first application setting. Its co-founders are Chen Wei (陈伟), formerly Li Auto’s (理想) AI chief scientist and foundation-model department head, and Zhang Xiao (张骁), formerly a Li Auto product-line president.

The stated uses of funding include embodied-model training, compute and data infrastructure, expansion of the core team, household-robot hardware research and development, and application validation. These are planned funding uses rather than disclosed spending already completed.

The company reported preliminary completion of its internally developed Ego data-collection equipment and data platform. It plans to establish a complete collection, cleaning, annotation, and training workflow during 2026, but the sources provide no compute-hardware inventory, infrastructure budget, financing conditions, or dilution terms.

• Xieyue Intelligence (斜跃智能)

🇺🇸Anthropic proposed continuous employee-level access for independent model evaluators; OpenAI committed to adopting the measure. #

Anthropic CEO Dario Amodei proposed that leading AI companies give embedded third-party evaluation teams continuous access comparable to that available to employees. The proposal specifies an ongoing evaluation arrangement, rather than merely a one-time external review.

Amodei said slowing capability development should not mean stopping model training or technical progress, but allowing sufficient time for alignment, safeguards, and third-party verification. The reports say he cited AI self-improvement and a reported OpenAI–Hugging Face agent incident as reasons for concern; they do not provide independent technical evidence about that incident.

The broader proposal also called for frontier-AI companies in democratic countries to coordinate common safety standards. A separate message records his call for the United States and other democratic governments to seek coordination with others.

Altman endorsed employee-like access for independent evaluators and said OpenAI would implement it, with more information to follow; Elon Musk also publicly agreed with Amodei’s position. These are proposals and commitments, not evidence that evaluator teams had already received access or that companies had implemented a coordinated training slowdown.

• Anthropic
• OpenAI

Server supplier iSoftStone Computing (软通计算机) won a RMB418 million award in China Mobile’s (中国移动) 2026–2028 server procurement. #

iSoftStone Computing (软通计算机) won an award in China Mobile’s (中国移动) centralized procurement for component-renewal servers covering 2026–2028. The September 12 report attributes the announcement to iSoftStone (软通动力) and gives the award value as RMB418 million.

The disclosed procurement concerns servers, not a separately identified purchase of graphics processing units (GPUs), accelerator cards, racks, or data-center power capacity. The source does not identify server quantities, processor architectures, accelerator configurations, or individual equipment prices.

The reported milestone is a procurement award, rather than completed delivery or recognized revenue. Delivery schedules, acceptance requirements, payment terms, and the award’s share of the overall procurement are not supplied.

• RMB418 million
• China Mobile’s (中国移动)

ZTE (中兴) shortened chip-adaptation cycles to 3–6 months with its domestic high-performance Matrix cluster supernode. #

ZTE’s (中兴) domestic high-performance Matrix cluster supernode was highlighted among 15 annual major breakthrough achievements announced at the 2026 China Computing Conference (中国算力大会). The report identifies the system as based on an OEX architecture, but does not expand that acronym.

The reported result is a reduction of the chip-adaptation cycle to three to six months. This is an integration-timeline metric, not a measurement of chip performance, cluster throughput, or model-training duration.

The source does not disclose the earlier adaptation baseline, supported chip models, accelerator count, or supernode configuration. It also supplies no benchmark methodology or independent test results for the reported reduction.

• ZTE (中兴)

Yangtze Optical Fibre and Cable (长飞) achieved 0.04 dB/km hollow-core-fiber attenuation and completed over 13 commercial and pilot projects. #

Yangtze Optical Fibre and Cable’s (长飞) hollow-core-fiber compute-interconnection technology was highlighted among the conference’s annual major breakthrough achievements. The report gives a minimum attenuation of 0.04 decibels per kilometer, a fiber transmission-loss metric rather than a compute-performance figure.

A separate manufacturing metric was a longest fiber draw of 91.2 kilometers. The report describes both the attenuation and drawing-length achievements as world records, but provides no underlying measurement documentation or independent record verification.

The company had completed more than 13 commercial and pilot projects in China and overseas. Because that count combines commercial deployments with pilots, it should not be presented as more than 13 fully commercial networks; customer names, individual route lengths, and project revenues are not disclosed.

Optical-module supplier Eoptolink (新易盛) increased 1.6T module shipments in Q2 versus Q1; 800G products became shipment mainstays. #

Eoptolink (新易盛), identified in the source by stock code 300502, said second-quarter shipments of its 1.6T optical modules increased markedly from the first quarter. It also reported continued growth in 800G shipments, with those products becoming shipment mainstays; the 1.6T and 800G labels refer to optical-module transmission rates, not compute capacity.

At its September 11 results briefing, the company described both 1.6T and 800G modules as principal delivery products for 2026. It reported a rising share of silicon-photonics products and progressing customer collaboration on linear pluggable optics (LPO), without providing shipment counts or percentage shares.

For near-packaged optics (NPO), management said customer collaboration was progressing and anticipated batch-delivery demand in 2027, followed by larger-scale demand and supply in 2028. It explicitly tied the ramp to market and customer demand, so those dates are management expectations rather than disclosed unconditional purchase commitments.

Management also expected a faster 1.6T ramp in the second half and continued silicon-photonics growth next year. Its positive market assessment was based on customer guidance and orders, but the supplied messages contain neither an order-book value nor a direct exchange filing supporting the outlook.

• Eoptolink (新易盛)

Photonic quantum-computer developer TuringQ (图灵量子) launched Gen3, a rack-mounted system incorporating programmable thin-film lithium-niobate photonic quantum chips. #

TuringQ (图灵量子) founder Jin Xianmin (金贤敏) unveiled TuringQ Gen3 at the September 12 Pujiang Innovation Forum (浦江创新论坛). The report describes it as the world’s first rack-mounted, chip-level scalable photonic quantum computer; that characterization is a reported launch claim, not independently established by the supplied material.

Gen3 incorporates high-speed programmable thin-film lithium-niobate photonic quantum chips. Its integrated modules include quantum light sources, programmable photonic quantum processing chips, single-photon detection, and heterogeneous quantum–classical computing.

The stated architecture supports expansion toward ten-thousand-photon-level resources on a single chip and one million modes. The report also states that chip-level quantum advantage was verified and that world-model applications would be explored, but supplies no benchmark details; photons and modes should not be relabeled as qubits or GPU equivalents.

Separately, TuringQ reportedly obtained registration for listing guidance and intends to pursue an A-share IPO, with Guotai Haitong (国泰海通) as its adviser. This is a preparatory listing-stage development, not an approved offering or completed flotation, and no fundraising amount, valuation, or dilution terms are provided.

• TuringQ (图灵量子)
• Gen3

China Mobile (中国移动) disclosed over 7 gigawatts of contracted AI data-center capacity, distinct from operational capacity. #

China Mobile (中国移动) general manager Chen Yangfan (陈扬帆) disclosed more than 7 gigawatts of total signed AI data-center capacity at the September 12 conference. The source uses a contracted-capacity description: it does not establish that this entire power capacity is built, energized, or operating.

Chen described plans for gigawatt-scale data campuses aligned with the Eastern Data, Western Computing initiative (东数西算). The stated approach includes coordination between computing and electricity resources, alongside cross-regional compute operations.

The company also outlined plans for multiple ten-thousand-card AI computing centers, inference clusters, and a distributed inference network. The source does not specify the accelerator types or establish that these planned card-scale centers are all deployed within the reported 7-gigawatt contracted total.

On networking, the report describes a deployed 400G optical transport network (OTN) backbone and an intelligent internet identified as G-SRv6, claiming world-first and world-largest status respectively. It also lists three latency service tiers—1 millisecond metropolitan, 5 milliseconds provincial, and 20 milliseconds nationwide—but gives no contract-level power allocations, construction schedules, capital expenditure, or customer concentration.

• China Mobile (中国移动)

China’s new-type energy storage (新型储能) reached 153 gigawatts of operational power capacity and 396 gigawatt-hours of energy capacity by June-end. #

China’s completed and operational new-type energy storage (新型储能) reached 153 gigawatts of power capacity and 396 gigawatt-hours of energy capacity by the end of June. The report also gives 61% year-over-year growth, without separately specifying growth rates for the power and energy measures.

Equivalent utilization hours increased from 611 hours in 2023 to 911 hours in 2024 and 1,195 hours in 2025. These are reported annual utilization metrics, not the discharge duration of an individual storage installation.

Liu Yongdong (刘永东), deputy secretary-general of the China Electricity Council (中国电力企业联合会), said installed new-type storage capacity grew more than fortyfold during the 14th Five-Year Plan (十四五). He described that period as a transition from early commercialization to development at scale.

For the 15th Five-Year Plan (十五五), Liu projected approximately 160 gigawatts of additional installed capacity. The cited planning figures target 300 gigawatts nationally by 2030, including 140 gigawatts of independent grid-side storage capable of supporting peak supply; these remain projections and targets, separate from June’s operational totals.

Two salt-cavern compressed-air energy-storage units in Jintan (金坛) completed full-unit startup, each with 350 megawatts of capacity. #

Two units at a salt-cavern compressed-air energy-storage project in Jintan (金坛), Jiangsu (江苏), completed full-unit startup on September 12. Each unit has a reported capacity of 350 megawatts, identifying a power rating rather than stored energy in megawatt-hours.

The report describes the project as the world’s largest salt-cavern compressed-air energy-storage project by total installed capacity. It characterizes the startup as progress in large-capacity technology and system integration, but does not provide independent comparative documentation.

Full-unit startup is the specific milestone disclosed, not a stated completion of commercial-operation acceptance. The source contains no energy-storage capacity, discharge duration, round-trip efficiency, project cost, or commercial commissioning date.

• Two

All 36 operating data centers in Langfang (廊坊) have begun transitioning into AI computing centers. #

All 36 operating data centers in Langfang (廊坊) had begun transitioning toward AI computing centers, according to the supplied report. The wording describes an underway conversion process, not 36 completed AI-center conversions.

The update accompanied the September 12 activation of the National Compute Interconnection Regional Node for Hebei (国家算力互联互通区域（河北）节点). This is a regional interconnection development, distinct from the physical conversion status of individual Langfang facilities.

The China Academy of Information and Communications Technology (中国信息通信研究院) also released its 2026 Comprehensive Computing Power Panorama Analysis (2026 综合算力全景分析), reporting that Hebei’s (河北) comprehensive compute index had ranked near the top for three consecutive years. No facility-level accelerator counts, megawatts, conversion budgets, or completion schedules are included.

China’s National Supercomputing Internet (国家超算互联网) aggregates over 3.5 million CPU cores and over 250,000 GPU cards. #

The National Supercomputing Internet (国家超算互联网) had aggregated more than 3.5 million central processing unit (CPU) cores and more than 250,000 GPU cards, according to Sugon (中科曙光) vice president Cao Zhennan (曹振南). The report also states that domestically sourced computing accounted for more than 95%, without specifying the denominator or calculation methodology for that share.

Registered users exceeded 1.7 million, while cumulative completed jobs exceeded 270 million. Daily jobs exceeded 300,000 on average and more than one million at peak, providing usage figures separate from the platform’s hardware inventory.

The application marketplace hosted more than 880 merchants and had recorded more than 240,000 orders across scientific research, AI, and industrial simulation. The aggregated cores and cards are platform-wide resources, not a disclosed configuration for one tightly coupled cluster, and the report does not identify GPU models or utilization by hardware type.

• National Supercomputing Internet (国家超算互联网)

China Computing Platform (中国算力平台) achieved nationwide integrated compute monitoring, connecting platforms across 31 provincial-level regions. #

China Computing Platform (中国算力平台) reported nationwide integrated compute coordination and monitoring, expanding from ten connected provincial-level platforms at the 2025 conference to platforms across 31 provincial-level regions. This is a monitoring and interconnection milestone, not a disclosed measure of newly constructed physical computing capacity.

Its compute marketplace had more than 10,000 registered enterprise users, over 200 compute-service providers, and more than 2,000 listed compute products. It also connected more than 300 large models and had accumulated billions of compute-monitoring data records.

A newly released small-business service list covers general-purpose computing, AI computing, large-model services, industry agents, infrastructure, storage, and security. Through the platform’s small-business marketplace section, enterprises can select container rentals and bare-metal resources or use model and agent services with token-based and elastic usage billing.

The policy context is the recently issued 15th Five-Year Plan for the Information and Communications Industry (信息通信行业发展“十五五”规划). The plan calls for a hub–regional–edge computing infrastructure hierarchy, automated monitoring, and improved supply–demand matching, but the messages provide no measured nationwide scheduling-success rate or utilization improvement.

• China Computing Platform (中国算力平台)

Xiongan International Integrated Compute Scheduling Center (雄安国际算力一体化调度中心) began operations, connecting 44 resource pools across 19 provinces. #

The Xiongan International Integrated Compute Scheduling Center (雄安国际算力一体化调度中心) announced the start of operations at the September 12 Xiongan forum. It comprises two modules: an integrated compute-scheduling platform and an international large-model service platform.

The scheduling platform covers eight national compute hubs and connects 44 compute-resource pools across 19 provinces. Its reported aggregate scale exceeds 27,000 “P,” but the source does not define the performance unit or numerical precision, so this figure cannot safely be converted into a specified floating-point benchmark or accelerator count.

The large-model service platform focuses on token scheduling and reportedly aggregates mainstream models across categories. Creating a globally oriented model-aggregation service hub is a stated objective; the source does not disclose international customer usage, transaction volumes, or the physical location of every connected resource.

China Mobile (中国移动) launched AI Trusted Computing (AI可信计算) for confidential cloud training, inference, and data use. #

China Mobile (中国移动) launched AI Trusted Computing (AI可信计算), abbreviated AITC, at the September 12 conference. The service is described as an integrated security offering spanning cloud-based AI training, inference, and the full lifecycle of data use.

AITC is built on Mobile Cloud’s (移动云) full-stack domestic heterogeneous-compute foundation. Its stated technologies combine confidential computing, domestic cryptography, and information-flow privacy protection, with offerings covering confidential general-purpose and AI computing as well as confidential tokens.

The announcement targets finance, government, industry, healthcare, and university research, identifying data leakage, model-asset theft, and lack of trust in compute environments as customer concerns. Its description of public-cloud use resembling private-cloud use is product positioning; no independent security audit, certification, performance overhead, pricing, or adoption figures are supplied.

• China Mobile (中国移动)
• AI Trusted Computing (AI可信计算)

UBTECH (优必选) began production at its Liuzhou (柳州) humanoid-robot factory, with planned annual capacity exceeding 10,000 robots. #

UBTECH (优必选) held the production-launch ceremony for its industrial humanoid-robot factory in Liuzhou (柳州) on September 12. The report states that a manufacturing approach using robots to build robots had entered mass production at the site.

The production line’s designed cycle permits one robot to come off the line every ten minutes, with planned annual capacity exceeding 10,000 robots. Both are design or planning figures, not disclosed actual annual output, sustained operating throughput, or customer deliveries.

The report describes the facility as UBTECH’s world-first industrial humanoid smart-manufacturing benchmark factory configured for ten-thousand-unit capacity. Liuzhou Party Secretary Zhang Zhuang (张壮) and UBTECH founder, chairman, and CEO Zhou Jian (周剑) attended, but no factory investment amount, utilization rate, order backlog, or product-model breakdown is provided.

• UBTECH (优必选)

World-model developer ShengShu Technology (生数科技) released Motus2, adding a lightweight tactile-expert module that refines manipulation actions using touch feedback. #

ShengShu Technology (生数科技) released Motus2 on September 12, describing it as a self-evolving general-purpose world model for dexterous manipulation. The specific technical update disclosed concerns contact during manipulation tasks.

Motus2 introduces a lightweight tactile-expert module that refines actions using the latest touch feedback. The stated function is to give action adjustments a contact-based input beyond vision, rather than relying solely on visual observations.

The source does not disclose model size, tactile-sensor configuration, training compute, task-success rates, or comparisons with earlier versions. It also provides no mechanism or benchmark supporting the broader “self-evolving” description, so no quantified manipulation improvement can be established from this announcement.

• ShengShu Technology (生数科技)
• Motus2

OpenAI Codex disabled an Astra context-management experiment affecting approximately 4,000–5,000 users, and removed misconfigured engines causing quality degradation. #

OpenAI Codex product lead Tibo reported several fixes following collaboration with users experiencing Astra quality problems. One issue involved skills written for earlier models triggering too frequently or preventing the model from checking its own work.

The team disabled an optional context-management experiment that could cause premature stopping or responses to older messages. Approximately 4,000–5,000 users were affected by that experiment, according to a rough estimate; this is not a disclosed count covering every reported Astra defect.

Separately, the team removed misconfigured engines that caused measurable quality degradation in the tail traffic routed through them. The source does not quantify that traffic share or the size of the quality decline, and it does not identify those engines as specific hardware devices.

Tibo also announced smaller improvements and a planned reset before local midnight, identified in the message as noon Beijing time. More consistent execution, better tracking of the latest user message, and improved checking of intermediate results were stated expected benefits, not independently measured post-fix outcomes.

• OpenAI Codex