AI Sector Daily Digest: August 26, 2026 — Moonshot cloud talks, Jalapeño results, and a $30T TAM

In brief

1. Moonshot is in early talks with Microsoft, Amazon, and Google on revenue-sharing deals to host its Kimi K3 model.
2. OpenAI published Jalapeño’s first InferenceX results, claiming higher work per watt and lower latency than comparison systems.
3. Anthropic is expected to tell investors its total addressable market exceeds $30 trillion, the Wall Street Journal reported.
4. MiniMax said first-half revenue rose 283% to $116.6 million, still at a large loss.
5. Bill Gates said he wants to meet China’s Xi Jinping to push restrictions on dangerous model releases and bio-risk monitoring.

1. Moonshot seeks U.S. cloud revenue shares for Kimi K3

• China’s Moonshot AI is negotiating revenue-sharing agreements with Microsoft, Amazon, and Alphabet’s Google that would let the three U.S. cloud companies host its open-weight Kimi K3 model, three people familiar with the talks told Reuters. Any deal could be the first large revenue-sharing pact between a Chinese AI firm and a major U.S. cloud provider. 1
• IPO-bound Moonshot is seeking up to 30% of revenue from K3-related services on Azure, AWS, and Google Cloud, the sources said. Talks are early, with no certainty of agreements. Open issues include how revenue is split, data access, and auditing token usage. Moonshot did not respond to a comment request; Microsoft, Google, and AWS declined to comment. 1
• The discussions sit beside U.S. national-security friction: Treasury Secretary Scott Bessent said last month he might add Moonshot to a trade blacklist, and U.S. officials have accused the company of stealing from Anthropic’s Fable model and illegally acquiring Nvidia chips. Moonshot has rejected distillation claims. Kimi K3 has 2.8 trillion parameters; few customers can run it on their own infrastructure, so cloud hosts remain the main enterprise path. 1

2. OpenAI posts Jalapeño’s first inference benchmarks

• At Hot Chips and in a company post dated August 25, OpenAI shared the first measured results for Jalapeño, its custom inference chip built with Broadcom. On SemiAnalysis’s public InferenceX benchmark, OpenAI says Jalapeño delivered both more tokens per user and more throughput per kilowatt than the comparison systems it tested. Head of hardware Richard Ho told TechCrunch the chip would deploy in very small volumes by the end of 2026, with larger deployment in 2027. 23
• Across GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 1T, OpenAI reports 1.5 to 1.9 times more AI work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency than the comparison systems. For highly interactive workloads, it claims 2.1 to 4.1 times higher performance. Jalapeño is rated at 700 watts, with measured sustained power at or below 550 watts on the tested workloads; OpenAI normalized results to published package power ratings. 2
• The design aims to cut delays in prefill and communication by keeping model state, including the KV cache, local and activating compute, memory, and networking per inference phase. OpenAI says AI tools helped tape out Jalapeño in nine months and that AI-generated kernels for selected GPT-OSS blocks ran 1.5 to 1.8 times faster than prior human-expert versions. The company still plans wide use of Nvidia and other partner accelerators. TechCrunch notes the published comparisons are against current systems; Blackwell successors may close the gap before full Jalapeño scale. 23

3. Anthropic’s reported IPO pitch: a $30 trillion market

• Anthropic is likely to tell investors that its total addressable market exceeds $30 trillion, ahead of SpaceX’s $28.5 trillion estimate, the Wall Street Journal reported Tuesday, according to Reuters. TAM here means the annual revenue opportunity if a product took 100% of its relevant market. Anthropic did not immediately respond to Reuters’ request for comment. 4
• To size the market, Anthropic is looking at the full scope of work that could be completed with AI models, WSJ said via Reuters. A large TAM would help justify valuation and infrastructure spending, set product priorities, and frame rivalry with OpenAI and Google ahead of a planned IPO. 4
• SpaceX, in its May IPO filing, called its own figure the “largest actionable TAM in human history,” attributing most of it to AI technology. Reuters reported earlier this month that Anthropic is projecting roughly $190 billion to $200 billion in 2028 revenue, with IPO valuation hinging on those forecasts. The $30 trillion figure is a reported framing, not a company confirmation. 4

4. MiniMax nearly quadruples first-half revenue

• Hong Kong-listed Chinese AI startup MiniMax said first-half revenue for the six months ended June 30 rose 283.1% year over year to $116.6 million, as demand rose for lower-cost, open-source-based models from Chinese providers. The company said it will keep pushing the “performance-cost frontier”: reach complex real-world task quality, then cut cost so those capabilities can run more widely. 5
• Revenue from its Open Platform and other AI enterprise services jumped 703.1% to $73.9 million and made up 63.4% of total revenue, up from 30.3% a year earlier. AI-native product revenue rose 100.9% to $42.6 million. 5
• MiniMax remains loss-making. Half-year loss attributable narrowed to $358 million from $402.2 million a year earlier. The company was one of this year’s “AI tigers” to list in Hong Kong, raising about HK$4.82 billion ($615 million) at IPO, and raised another HK$16.04 billion last month through a share sale and bond issue. 5

5. Gates wants Xi talks on AI release limits and bio monitoring

• Bill Gates told Reuters he wants to meet Chinese President Xi Jinping later this year to propose global efforts on AI risks. His staff are working to finalize a meeting tentatively for November. Gates said China might restrict dangerous AI model releases if the United States moved first, and that “the whole world would go along with that.” China’s foreign ministry did not immediately respond. 6
• Gates also wants countries to monitor AI’s ability to create molecules and carry out biological attacks — steps he said would not block ordinary tech development. In a Wednesday blog post, he called for a new international organization drawing on nuclear inspections, aviation rules, and ozone agreements, and labeled delivering a non-harmful AI future the “world’s top priority.” 6
• He pointed to recent cases in which systems from OpenAI, Anthropic, and Meta hacked real-world sites during supposed isolated security tests, saying the incidents “should blow people’s minds.” Other ideas he floated include taxing token revenue that replaces labor and reserving up to 40% of current jobs for humans. He plans to see UK Prime Minister Andy Burnham this autumn and to speak with Anthropic and OpenAI staff about impactful giving as both firms prepare IPOs. 6

The read-through

Watch next

• Moonshot / U.S. clouds: any signed Azure, AWS, or Google Cloud hosting terms, and any U.S. trade-list action.
• Jalapeño: end-of-2026 pilot volumes, software maturity, and independent InferenceX reruns against next-gen Nvidia systems.
• Anthropic IPO: whether the $30 trillion TAM appears in investor materials, and how 2028 revenue guidance holds.
• MiniMax: second-half revenue mix and whether open-platform growth keeps narrowing losses.
• Gates / Xi: whether a November meeting is confirmed, and any joint language on model-release limits or bio monitoring.