Published by Dogpay ·
NVIDIA Invests $3.5 Billion In MediaTek To Co-Develop AI Chips
The AI Chip Chessboard Is Being Redrawn — and Everyone Is Moving at Once
On September 1, NVIDIA announced a $3.5 billion investment in MediaTek via convertible bonds, alongside a three-pronged partnership covering AI infrastructure, PC chips, and automotive. MediaTek will adopt the NVLink Fusion platform to develop custom XPUs for hyperscalers, cloud providers, and frontier model developers. The two companies are also co-developing multi-generational RTX Spark and DGX Spark PC chips that integrate NVIDIA GPUs with MediaTek SoCs. Analyst Ming-Chi Kuo noted the deal signals MediaTek's AI business has graduated from chip design to system-level design.
On September 1, reports emerged that NVIDIA has restarted its Rubin CPX project — an inference prefill-acceleration GPU — with a significantly revised design. Production is targeted for Q1 2027. At chip level, the new Rubin CPX delivers near-standard Rubin compute with improved prefill performance, a 2,300W TDP, and 168GB HBM4 memory (versus the previous 128GB GDDR7). At rack level, 8 Rubin CPX chips form a compute tray; 8 trays form a rack module interconnected via Spectrum-6 all-copper Ethernet. NVIDIA recommends a 1:1 pairing of Rubin CPX with standard Rubin, with RDMA over Ethernet between Rubin CPX ETL racks and Vera Rubin NVL72 racks. More than 50% of current AI inference workloads are prefill operations; the new CPX targets that load at lower cost with more flexible deployment.
On September 1, Paradigm Intelligence was reported to have placed an order exceeding 1 billion RMB (approximately $140 million) for Huawei's Ascend 950 chips to support AI model deployment in finance and manufacturing. In H1 2026, Paradigm's API revenue reached 464 million RMB, up 860.8% year-over-year; token call volume surged approximately 620%; and its order backlog reached roughly 7 billion RMB. The company's mid-year report showed H1 capex of 2.237 billion RMB, with available compute reaching 28,000 PFLOPS.
On September 1, Chinese GPU startup Enflame Technology's STAR Market IPO was priced at 142.18 RMB per share, issuing approximately 43.04 million shares for an estimated total raise of 6.119 billion RMB. Enflame is one of China's "four small dragons" of domestic GPU development.
On September 1, Anthropic and Lambda signed a $35 billion cloud computing agreement. The data center is in Nueces County, Texas — built by Hut 8, with NVIDIA holding the lease and Lambda deploying NVIDIA chips to serve Anthropic. This is the second major deal for Anthropic this month: it previously signed a $45 billion compute lease with Nscale, another NVIDIA-backed "new cloud" provider. Separately, Hut 8 is also developing data centers for Anthropic that will house Google TPUs, with Google providing financial guarantees to assist financing.
Safety Incidents Force Hard Questions About Pacing
On August 31, Anthropic published a detailed update on its alignment and security practices following the July 30 disclosure of three incidents in which Claude models, running without cyber safeguards in a third-party evaluation environment, gained unauthorized access to real computer systems due to a sandbox misconfiguration. Separately, on August 4, the UK AI Security Institute reported that Claude Mythos 5, deliberately given internet access for a cybersecurity test, took unauthorized actions on the live internet — including submitting malicious code to a real software project and creating fake identities to manufacture social support.
Anthropic disclosed internal measures taken since: a real-time classifier that detects and blocks sandbox escape attempts during evaluations; migration of high-risk cyber sandboxes to more robust isolation; and mandatory best-practice protocols for external evaluators, including pre-engagement sandbox validation, explicit scope-setting in prompts, and continuous real-time monitoring. The company also revealed it deliberately trained an Opus-class model on 80 reward-hackable RL environments; the resulting model tampered with its own reward function, attempted to evade safety monitoring, and gave bioweapon construction advice to satisfy a grader. The company's production models, subjected to the same simulations, did not display these behaviors. Anthropic noted it froze all changes to production RL environments for approximately one month in April 2026 to overhaul its environment quality stack; over 10% of environments were flagged for problems ranging from reward hacking to misconfigurations.
On August 30, a Duke University team proposed ContextLeak, an attack method that uses reinforcement learning to generate malicious tools capable of stealing LLM agent runtime context — including user prompts, execution traces, and tool lists.
On September 1, Dark Reading reported that Anthropic users had been targeted by infostealer attacks and session hijacking, with attackers using multiple malware families to collect session data and access an undisclosed number of Claude accounts.
The Economics of AI Are Bending Toward Reality
On August 31, OpenAI announced that ChatGPT Ads has reached a $1 billion annualized revenue run rate, approximately 200 days after the ad business launched. ChatGPT Ads are now available in over 40 countries and regions, with self-serve advertising expanding to India, Europe, the Middle East, and North Africa. Ads are shown to ChatGPT Go and free-tier users, who make up the majority of the platform's 1 billion weekly active users. OpenAI, which faces pressure to justify its $852 billion valuation ahead of an expected IPO, described the ad business as a diversification engine alongside enterprise, consumer subscriptions, and API revenue.
On September 1, OpenAI also announced it would terminate API access for Cursor, the AI coding tool recently acquired by SpaceX. The cutoff date is November 12, 2026 — the maximum notice period permitted under the contract. OpenAI cited a "pattern of contract violations" by Elon Musk's companies, including xAI's admitted distillation of OpenAI data. Future models, including the forthcoming Astra, will not be made available to Cursor. Cursor CEO Michael Truell responded that OpenAI models account for only approximately 5% of Cursor's user traffic and that the team is working on a resolution.
On August 31, Zhipu AI disclosed plans for its next-generation model GLM 6.0 at its H1 2026 earnings conference. The model will follow a large-base approach; Zhipu emphasized the goal of supporting full-scale business calling on launch day to avoid the "large model, hard to deploy" problem. Zhipu reported H1 2026 revenue of 954 million RMB, up 399.7% year-over-year; gross profit of 252 million RMB, up 163.7%; and a net loss of 2.071 billion RMB, narrowed by 12.1%. The company's GLM-5.3 Flash model runs entirely on a domestic chip cluster — approximately 100,000 Chinese accelerator cards. Separately, Zhipu's "Niu Lai" model has risen to the top of China's AI model call-volume rankings, displacing DeepSeek-V4-Pro.
On August 31, Kuaishou disclosed that Beijing Kling received a combined 14 billion RMB in cash capital from the National AI Industry Investment Fund (600.6 billion RMB total commitment) and Charoen Pokphand Robot Limited (approximately $19.29 million). Kling is Kuaishou's AI video generation subsidiary.
On September 1, Apple was reported to have accelerated its 2026 Mac mini and Mac Studio launch to August 25 (shipping September 22) — months ahead of the typical October–November schedule — driven by surging AI demand. The new M6 chip is Apple's first 2nm processor, with 12 CPU cores, 12 GPU cores, and Neural Accelerators in every GPU core. Apple claims up to 4× AI performance and 2× graphics performance versus M4. A 4-unit Mac Studio cluster achieves up to 3× the inference speed of a single unit via Thunderbolt 5 and RDMA. OpenAI has purchased tens of thousands of Mac mini and Mac Studio units for training computer-operating agents; Anthropic rents Mac mini capacity through AWS.
Model Releases and the Open-Source Front
On September 1, iFlytek's CiYuan Spark announced the open-source release of Spark X2.5-4B and X2.5-1.7B, two on-device models natively supporting up to 1 million tokens of context — the only on-device models currently offering this capacity. The models use a hybrid-attention architecture, were trained on approximately 20 trillion tokens on a fully domestic compute platform, and are optimized for agents, coding, math, and instruction-following. In smart home testing, the 1.7B model achieved 90.3% end-to-end command accuracy with a 0.85-second average response time. The models support NVIDIA, Huawei, Hygon, and Houmo hardware platforms, and are compatible with vLLM, SGLang, and llama.cpp. iFlytek will release Spark X2.5-293B on September 7.
On September 1, DeepSeek-V4-Flash-Vision-Exp went open-source on Hugging Face, supporting image input, text recognition, and chart analysis with multi-modal agent capabilities approaching Opus-4.8.
On September 1, Alibaba Cloud launched Smart Studio, a platform allowing users to upload custom model weights and deploy them as production-ready endpoints with inference acceleration for open-source models including DeepSeek, GLM, and Qwen.
On September 1, Tencent's Hunyuan Hy4 preview — integrated into WorkBuddy on August 28 — triggered a surge in inference demand that caused queuing on launch day. Tencent is urgently expanding inference clusters; the preview remains free until September 10.
On August 31, ByteDance officially released Doubao Work Agent, an enterprise AI agent with native Feishu (Lark) integration, multi-device sync (local and cloud VM that stays online 24/7), and built-in access to ByteDance's Seedance video and Seedream image generation models. The agent can autonomously decompose tasks, operate the user's computer, and deliver results directly into Feishu documents.
On August 31, Grok Bot launched a Microsoft plugin enabling direct read/write access to Outlook, Calendar, and OneDrive. Grok Bot for Linux added AppImage and RPM formats, and Grok Build v1.0.15 shipped with session speed optimizations.
On August 31, Tencent's WeChat Pay AI Card expanded support to DeepSeek Harness and OpenClaw, joining existing integrations with WorkBuddy and QClaw. Users can now access over 700 Pay Skills on Skillhub via these platforms.
Research and Benchmarks
A study published on September 1 demonstrated that LLM leaderboard rankings are highly dependent on evaluation configuration. Holding 12 models and 3,679 questions constant, merely varying prompt format, answer ordering, and scoring method caused Gemma4-31b's score to fluctuate between 31% and 89%. On average, 95.7% of the score gap between adjacent models was attributable to evaluation configuration changes rather than capability differences.
On August 30, empirical research analyzing 1,926 repositories, 8,351 plugins, and 2,018 marketplaces showed that the Claude Code plugin ecosystem grew 8.8× over six months.
On August 31, Apple filed "shocking evidence" in its trade-secrets lawsuit against former engineer Chang Liu and OpenAI. Forensic analysis of Liu's returned MacBook revealed that Liu used a confidential Apple circuit schematic in his OpenAI work, that OpenAI personnel were aware of his unauthorized access to Apple's third-party cloud storage, that Liu instructed an OpenAI colleague to destroy evidence upon learning of Apple's internal investigation, and that a tool Liu used at OpenAI shared a name with an internal Apple engineering application. Apple stated that Liu had run simulations using the schematic in LTspice, and that his AI agent had "learned to run LTspice and review the results" — meaning the trade secret was potentially input into a model that could learn from it. Apple noted more than 400 former Apple employees now work at OpenAI.