阿里Qwen3.8-Max低价开放扩散，阿里云叙事利好但港股窗口已消化

# 美股盘后催化 6/12: 中国开放模型Qwen/Kimi逼近前沿且推理成本更低，阿里AI云受益但模型价格战升级

- 美股会话: 2026-08-03
- 方向: MIXED；类型: 产品/技术；规划重大性: 8/10
- 映射模式: competitive
- 美股驱动（规划器初判，需独立核验）: 事实：多条材料称阿里Qwen新旗舰模型低价开放API、计划开放下载，Kimi K3和中国开放模型在成本与能力上接近美国前沿模型；阿里BABA上涨。推断：对阿里云AI生态、开发者采用和中国企业AI应用是正面，但对基础模型收费、闭源模型溢价和算力投资回报构成压力。

## 本催化相关美股真实收盘行情

- 英伟达 (NVDA.US): 206.64 +2.93%
- 微软 (MSFT.US): 487.65 +4.93%
- 谷歌 (GOOGL.US): 373.51 +4.88%
- 亚马逊 (AMZN.US): 284.02 +4.58%
- 阿里巴巴 (BABA.US): 127.3 +4.13%

## 亚太映射假设（仅候选，必须查证）

- 阿里巴巴-W (9988.HK) [HK/同源本地股/positive]: Qwen性价比和开放生态若带来API调用、企业客户和云资源消耗增长，可通过ADR/H股平价和云估值传导。；核验问题: 核实Qwen版本、API定价、调用量、企业付费转化和阿里云AI收入贡献。
- 腾讯控股 (0700.HK) [HK/竞争替代/mixed]: 低成本开源模型可降低应用门槛，但压缩模型层收费空间并加剧国内云竞争。；核验问题: 核实腾讯混元、云AI定价和企业客户采用是否受Qwen/Kimi影响。
- 百度集团-SW (9888.HK) [HK/竞争替代/mixed]: 文心API和百度云可能面对低价开放模型竞争，同时受益于企业AI需求扩大。；核验问题: 核实百度文心价格、调用量、客户流失或新增情况。

## 必答问题

- Qwen/Kimi真实企业付费调用量和毛利率如何？
- 高效推理模型会降低总GPU需求，还是通过价格弹性扩大token消费？
- 开源版本许可、参数和商用限制是什么？

## 原始证据全文（7 条）

### [E0230] The Information | 2026-08-03 14:01 UTC | 盘中
### AI ‘Air Pocket’

While major AI spenders including Google are showing accelerating sales of either software or cloud services, Google shareholders have grown a little jittery about the company’s rising cash burn from spending around $20
URL: https://www.theinformation.com/articles/google-deepmind-exec-says-unprecedented-capex-actually-bet-rsi
### AI ‘Air Pocket’

While major AI spenders including Google are showing accelerating sales of either software or cloud services, Google shareholders have grown a little jittery about the company’s rising cash burn from spending around $200 billion *this year* on AI data centers and other equipment, and its plans to spend significantly more next year.

The revenues from AI “don’t sustain the capital expenditures we’re making so far,” said Sekhon, who joined Google in April from investment firm **Bridgewater Associates** and was previously a statistics professor at **Yale University**. That means “we have this danger we could hit an AI air pocket such that the expenditures happen but the revenues don’t show up.”

Sekhon did not mince words in saying that the industry’s capex spending is the “biggest scientific bet civilization has ever made,” dwarfing the U.S. government’s Apollo missions, the Manhattan Project, and spending to develop the internet. (The railroad buildout may have cost humanity more than AI, but “we already knew how to build them” so it didn’t qualify as a scientific bet, he said.)

But while the industry’s current technology doesn’t qualify as RSI, “betting against it would appear to be unwise,” he said. Today we can see “precursors” of RSI as AI firms use models to help design parts of other models. However, Sekhon said “this should not be shocking because steam engines were used to create the next steam engine.”

### True RSI

I wouldn’t blame readers for wondering whether RSI is just rhetorical nonsense, the way AGI is turning out to be. But nobody can deny that the industry has come a long way in just a few years: from a fixed, static model you can talk to, but which has no fresh information or an ability to act; to one that could go look for information on the internet; reason through complex problems by using up more computing resources before giving an answer; read and write code across many programming languages; perform actions such as using a browser, a computer or any software applications.

Today’s AI agents also can tap data from many sources to solve tasks; record the actions they took so developers can later diagnose mistakes; read and learn from human instructions on how to perform certain tasks; recover from mistakes and continue working on problems for hours with limited human intervention; and spawn “subagents” to handle smaller pieces of a bigger problem.

Notably, AI can also generate or improve low-level software, such as kernels and compiler components, to help AI models run more efficiently on specialized servers. And researchers can use a model’s own answers as training data to update its weights—the numerical settings that shape how the model behaves—and improve its future responses.

True RSI would go much further, repeatedly improving the software and training processes that enable it to make even more improvements, creating a kind of loop. One researcher put it this way: RSI means that a model could independently redesign its entire architecture and develop an entirely new model.

We are nowhere near that stage, nor are we close to an AI that can come up with brand-new scientific discoveries, for instance.

But based on recent progress, Sekhon said he believed RSI would “probably [arise] within the next couple of years.”

Later at the event, Sekhon’s DeepMind colleague **Oriol Vinyals** and **OpenAI** co-founder **Wojiech Zaremba** said during a panel that they believed RSI might be achieved by 2027 or 2028.

### Biological Attack Risks

Of course, for all the benefits of RSI, there are also the risks. AI could cure diseases and help humanity “explore the cosmos,” Sekhon said. But he and **Dawn Song**, a Berkeley computer science professor who just joined **Meta’s “Superintelligence”**unit, spent a lot of time at the event discussing risks from AI, especially as a tool for cyber and biological attacks.

AI will “benefit attackers more in the near term,” Song said, because of the asymmetry in which “attackers only need one successful exploit” while “defenders have to defend against all attacks.”

Sekhon agreed, saying that as attackers poison open source repositories and exploit poorly written human code, “it’s going to be rough for a while. We have a lot of weak systems, like the United States energy grid, hospitals and the list goes on.”

Eventually, defenders will be able to utilize AI to fend off AI-based attacks and more companies and institutions will create code for their systems that will (hopefully) be verifiably free of vulnerabilities.

But cybersecurity might seem like child’s play compared to biological attacks stemming from AI, Sekhon suggested.

“We are quite exposed to bio risk as it is,” he said. “We are very close to a world where someone can design a virus or a protein just by talking to a model in natural language. That’s a very dangerous world that is attacker-advantaged even in the long run.”

Defending against such threats will require much more licensing, monitoring and tracking of “biologically relevant materials,” similar to the way society now tracks fertilizers that can be used for explosives, he said.

He also said Google is working to [adapt its watermarking technology (SynthID), which identifies AI-generated content, for biology,](https://deepmind.google/blog/our-approach-to-bioresilience/) he said. That could help companies that synthesize DNA for pharma or biotech companies to screen for potentially risky, AI-generated biological sequences, for instance.

*Here’s what else is going on…*

### Big Number

**Drumbeat Capital**, which invests in robotics, semiconductors and other deep tech industries, released [a report](https://drumbeat.capital/) on the global state of deep tech. The report found that global deep tech investment in 2025 reached $183 billion, up 58% from its previous peak in 2021, while other tech funding is down about the same proportion from 2021.

### Overheard

**Anthropic** [said its AI models hacked](https://www.theinformation.com/briefings/anthropic-says-models-hacked-outside-sites-testing) into three outside organizations during testing of their cybersecurity capabilities, incidents it discovered during a review prompted by a similar episode at rival **OpenAI** that has fueled widespread concern about AI safety.

**Tesla** CEO **Elon Musk** [denied](https://www.theinformation.com/briefings/elon-musk-dismisses-report-tesla-considered-selling-chinese-business) a Wall Street Journal report that Tesla was considering selling off its business in China to prepare for a merger with **SpaceX**. Musk said the report was “fake news” in a post on X Friday morning.

### People on the Move

**Ori Herrnstadt**, an engineering veteran of both **Apple** and **Google**, [joined](https://www.theinformation.com/briefings/exclusive-aws-taps-apple-executive-lead-key-ai-products) **Amazon Web Services** in May as vice president of compute AI services, an AWS spokesperson confirmed.

**Coinbase** legal chief **Paul Grewal** is leaving to join AI coding startup **Cognition** as its general counsel.

### Earnings Report

**Amazon** stock [surged](https://www.theinformation.com/briefings/aws-revenue-growth-accelerates-sharply-37-q2) in after hours trading after the ecommerce-and-cloud firm reported that revenue from its cloud division **Amazon Web Services** soared 37% in the second quarter, nine percentage points faster than the first quarter, bringing the division’s revenue to $42.2 billion. AWS’ operating margin also improved to 39% compared with 37% in the first quarter.

**Apple** [expects choppy waters](https://www.theinformation.com/briefings/apple-expects-supply-constraints-hit-hard-current-quarter)ahead for the quarter that ends in September, the company said on its June quarter earnings call. Apple’s Chief Financial Officer **Kevan Parekh** warned that supply constraints for advanced chip manufacturing will “increase significantly.”

### Deals and Debuts

*See The Information’s [Generative AI Database](https://www.theinformation.com/projects/generative-ai) for an exclusive list of private companies and their investors.*

**Amazon** [has completed its $50 billion investment](https://www.theinformation.com/briefings/amazon-completes-additional-35-billion-investment-openai) in **OpenAI**, putting in the remaining $35 billion in two stages in recent months, the commerce-and-cloud firm revealed in a securities filing on Friday.

**Xsight Labs**, a chip company that designs hardware to move data faster inside data centers, raised more than $300 million in a funding round led by **Fidelity Management & Research Company**.

**Onyx Security**, a company whose software lets enterprises discover, monitor and rein in the AI agents running inside their systems, raised $113 million in a Series B funding round led by **Bessemer Venture Partners**.

**DataBahn**, a startup whose software sits between a company's many data sources and the tools that consume that data, raised $40 million in a Series B funding round led by **Insight Partners.**

**Okta**, the identity-management company, signed a definitive agreement to acquire **Permiso Security**, an identity-security startup that detects and mitigates threats across humans.

**OpenAI**is [preparing to release a new model family](https://www.theinformation.com/briefings/exclusive-openai-previews-astra-ai-model-dc), tentatively using the name **“Astra,”** with improved abilities to complete long-running tasks, The Information [reported](https://www.theinformation.com/briefings/exclusive-openai-previews-astra-ai-model-dc).

[Chinese AI developer DeepSeek’s release](https://www.theinformation.com/briefings/deepseek-makes-splash-small-affordable-v4-flash-model) on Friday of a smaller version of its flagship **V4** open-source model is causing a stir. In early tests, the model, **V4-Flash**, demonstrated capabilities comparable to much larger open-source and proprietary models, and it’s also less expensive than those rivals.

Chinese tech giant **Alibaba Group** [made](https://www.theinformation.com/briefings/alibaba-offers-new-flagship-model-lower-prices-kimi-k3) its new flagship model, **Qwen3.8-Max**, widely available on Monday through application programming interfaces at lower prices than its Chinese competitor **Moonshot AI’s** popular **Kimi K3** model. Alibaba will also make the open-source model available for download next week.

Chinese AI firm **MiniMax** on Friday [announced a new open-source video generation model](https://www.theinformation.com/briefings/chinas-minimax-launches-new-open-source-ai-video-model), **H3**, cranking up competition against AI video rivals like **ByteDance** and **Google**.

**OpenAI** [said it is cutting prices](https://www.theinformation.com/briefings/openai-slashes-prices-newest-models) on two of its newest models weeks after their release, as AI companies respond to customer concerns about surging bills for their services.

### New From Our Reporters

[https://www.theinformation.com/articles/apple-icloud-policy-fueled-employee-leaks-ahead-openai-suit](https://www.theinformation.com/articles/apple-icloud-policy-fueled-employee-leaks-ahead-openai-suit)Exclusive

#### [How an Apple iCloud Policy Fueled Employee Leaks Ahead of OpenAI Suit](https://www.theinformation.com/articles/apple-icloud-policy-fueled-employee-leaks-ahead-openai-suit)

By Aaron Tilley

[https://www.theinformation.com/articles/robinhood-now-makes-revenue-predictions-stock-trades](https://www.theinformation.com/articles/robinhood-now-makes-revenue-predictions-stock-trades)Sunday Insights

#### [Robinhood Now Makes More Revenue From Predictions Than From Stock Trades](https://www.theinformation.com/articles/robinhood-now-makes-revenue-predictions-stock-trades)

By Yueqi Yang

[https://www.theinformation.com/articles/garry-tans-political-dreams-require-32-year-olds-grunt-force](https://www.theinformation.com/articles/garry-tans-political-dreams-require-32-year-olds-grunt-force)#### [Garry Tan’s Political Dreams Require This 32-Year-Old’s ‘Grunt’ Force](https://www.theinformation.com/articles/garry-tans-political-dreams-require-32-year-olds-grunt-force)

By Eli Rosenberg

## Recommended Newsletter

Stay ahead of the AI infrastructure wars. Ann Davis Vaughan covers how tech titans are disrupting the energy sector and becoming accidental industrial giants to compete on the AI model battlefield. Subscribe to [AI Infrastructure](https://www.theinformation.com/features/ai-infrastructure?rc=zjctrx).

## What We’re Reading

#### [Situational Awareness got the future right but misread the past](https://www.ft.com/content/20af5836-51b0-4a37-a820-9de1583a18b4?syn-25a6b1a6=1)

#### [Jefferies Analyst Calls Out ‘Simple-Minded’ Criticism of SpaceX’s Governance](https://www.bloomberg.com/news/articles/2026-08-02/jefferies-analyst-slams-simple-minded-criticism-of-spacex-governance?srnd=phx-technology)

#### [This CPO regrets that product management exists | Tom Verrilli (CPO of Whatnot)](https://www.lennysnewsletter.com/p/this-cpo-regrets-that-product-management)

### Upcoming Events

#### [Wednesday, September 23 — AI Agenda Live SF 2026](https://www.theinformation.com/events/ai-agenda-live-2026)

Join us to explore the real-world shifts, new architectures, and enterprise strategies defining the next era of AI. Secure your spot today to hear directly from top minds at Google DeepMind, OpenAI, Atlassian, and more.

[More details](https://www.theinformation.com/events/ai-agenda-live-2026)

#### [Tuesday, October 27 — The Information’s 2026 WTF Summit](https://www.theinformation.com/events/wtf-summit-2026)

Tickets on sale now: The Information returns to Napa Valley October 27–28 to convene senior women across tech, media, and finance. Join us for two days of intimate, candid conversations with the leaders navigating today’s global shifts.

[More details](https://www.theinformation.com/events/wtf-summit-2026)

*Thank you for reading the AI Agenda Newsletter! I’d love your feedback, ideas and tips: [stephanie@theinformation.com](mailto:stephanie@theinformation.com).*

*If you think someone else might enjoy this newsletter, please pass it forward or they can sign up [here](https://www.theinformation.com/newsletters/ai-agenda).*

### [E0299] Qwen | 2026-08-03 14:09 UTC | 盘中
Ask anything, anonymously. Qwen3.8-Max has landed on Venice. Give it a try!
URL: https://x.com/Alibaba_Qwen/status/2084280439909589230
Ask anything, anonymously. Qwen3.8-Max has landed on Venice. Give it a try!

### [E0852] Bloomberg TV+ | 2026-08-03 15:30 UTC | 盘中
「Bloomberg TV+」直播转录 23:00–23:30
URL: http://39.105.42.197:20307/event/live
【实时直播转录 · Bloomberg TV+】
来源:Bloomberg TV+(bloomberg 直播间 btv_us) · 标题:Bloomberg TV+
时段:2026-08-03 23:00–23:30(Asia/Shanghai),共 189 段
说明:以下为该时段该直播间的连续语音转录(直播 ASR,可能有同音/识别误差)。请从中识别与 AI算力/芯片/服务器/光模块/液冷/数据中心/大模型/机器人·具身智能/自动驾驶 等相关的、**可交易**的公司/产品/订单/产能/需求/合作/高管或官员表态,判断是否构成可交易事件;若通篇无可交易内容,判为不做。识别误差请结合上下文与常识修正,勿据孤立音近词臆断。

--- 转录正文 ---
[23:00] US models, the frontier labs models add some benchmarks, so for example, on research reproduction on agent use of computer.
[23:00] These are strong suits. It also says that its model has 2.4 trillion in terms of parameters, and that's basically on power.
[23:00] the level that we've been seeing for the Kmi K3 model, which was out a few weeks ago. Also the company yes also said.
[23:00] Something about cost, obviously per token cost, it's a bit lower than the US frontier models as well,$2 per$1 million tokens.
[23:00] The input side, you have a really strong command of Chinese technology companies from the stock perspective, right, Is Alibaba a name Heri?
[23:00] Ts on its AI prowess, or like some of these American multifaceted companies, a different part of its business.😊
[23:00] Yeah, the AI story is really important for Alibaba, if we zoom out a bit, today it gains about 5%.
[23:00] If you really look at the trajectory since June, you will see that the stock is up about 35,40% from a June low and.
[23:01] The key story here is about the more optimism in AI, obviously we're seeing thequin break through today but.
[23:01] Last month the company has a closed store meeting with analysts and analysts came out more optimistic about the company's cloud revenue growth.
[23:01] For example, JP Morgan is modeling about 45% in revenue growth for its cloud unit for the June quarter, so there has been an acceleration over there.
[23:01] But still on the main business, on the e-commerce business, because it has been a company known for its main shopping site.
[23:01] In China, that part of business is still lacking, although it's still the cash cow so to speak, so investors' opinions are still mixed.
[23:01] That's why we are seeing sentiment has been pretty weak into June, but since June, some positive AI announcements has really been boosting the stock.
[23:01] Reren you very Alibaba says its new Quen model delivers frontier AI performance and as Henry outlined at a fraction.😊
[23:02] Of the cost could cheaper AI accelerate adoption rather than just pressure pricing. joiningining us now is Lauren Cassidy.
[23:02] officer of the founderss 100 ETF, the ETF tracking founders in this space, Meder and Zuckerberg, Musk and space.
[23:02] The economics of AI is really what the Chinese story is putting front and center. So I pose that question to you, drive adoption.
[23:02] Or lower the tokenomics here Yes, I do see this as a positive.
[23:02] And that there's room in the world for both open and close models, and I do think it will drive prices a little lower.
[23:02] Isn't cheaper than anthropics, haiku, but if it is fable level performance, then it's about a fifth of the price.
[23:02] And so I think that that could put pressure on the L limbs, but over.
[23:02] There's still a lot that the American Ls have to offer What are the advantages that the U models have.
[23:03] Laauuren trust reliability, additional features, convenience of not.
[23:03] HavingTo modify the models so it just depends what an enterprise is looking for, what.
[23:03] It's the I'm looking at the found 100 ETF And again, feel free to remind us of the.
[23:03] On the founder side, but you don't have any of the ADRs for some of these big Chinese tech companies.
[23:03] The reasons for that I think that we believe that.
[23:03] In American innovation long term and just as we've had waves of innovation and.
[23:03] Greater adoption in the country of the steam engine, electricity, modern manufacturing computers.
[23:04] And the internet, we are now on the sixth wave, which is AI, we think it's going to be a really big thing we see.
[23:04] Tremendous adoption over the next decade, and we are very exposed to those stocks.
[23:04] Let's shift gears a little bit, SpaceX has its first quarterly earnings and called Tuesday E.
[23:04] ingThis is what I wrote published this morning. It's the first opportunity that Spacex investors might have.
[23:04] To put the question of a Tesla Spacex merger to Elon Musk, maybe you and I have discussed this before from the test.
[23:04] That's the side of the table. But where does founder 100 ETF stand on that idea.
[23:04] I believe that's inevitable over time, Elon Musk is an incredible founder and he is building.
[23:04] The only closed loop AI network, and over time there's going to be more and more opportunities for.
[23:05] Tsla and SpaceX to share they're going to be building Terraabab I'm really hoping to hear details about that.
[23:05] Build out tomorrow, they'll be sharing chips, there's going to be a lot of opportunities for synergies between the businesses.
[23:05] Probably makes sense at some point to fold Tesla into the greater mix as weve work towards data centers in space.
[23:05] Lauren Cassidy, a found a 100 ETF back on the show, thank you very much indeed, another top story, one of the crypto industry.😊
[23:05] most trusted storage devices is at the center of a major hack. Coinkite, the Canadian company behind the cold card.
[23:05] Hardware wallet says a software floor made some bit recovery keys predictable, allowing hackers to drain wallet.
[23:05] That was supposed to be kept safely online, According to Galaxy Research, more than 1700 Bitcoin worth about 10.
[23:06] has been stolen from roughly 5,000 different wallets. Coinkite says it has released a firmware update.
[23:06] To fix the floor, another stock that we're watching because of what happened over the weekend, shares of AMC now modestly high.😊
[23:06] 7 tenth0 of a percent have been much higher at the open, High ever total weekend revenue, including.
[23:06] For admissions and food and beverage, by the way, why the driven by the opening of Spiderman, brand new Day yet.
[23:06] To see it, that film grossing $355 million domestically in the Us. All right, Coming up. Valalour Atomics just secured.
[23:06] 🎼One billion in its series B funding round joining us next is Valaraatomic CEO and founder Isaiah Taylor.
[23:06] This is Flinberg Tech.
[23:07] 🎼The top names in global aviation are on bloomberg I to go for that getting worse I got an old price that said9 buck.
[23:07] 🎼So about any sign of this show that consumers are starting to get a little bit more cautious again, so very complex economies very complex.
[23:07] but we have more visibility than others with the backlog we have, we're ramping up production across many of our product lines, markets are very strong.
[23:08] So our backlog is very, very strong so our story is really focusing on executing that backlog we're on a multi year ramp both with our.
[23:08] And in defense right we sold out into the early 2030s we have our work cut out for on a multi year basis.
[23:08] 🎼Customers and governments with new means of facing those stress. and this means accelerating We need to expand.
[23:08] apacTo get after both speed, as well as integrated, interoperable allied capability, which will help all our nations feel safer.
[23:08] 🎼New that matters is on Bberg context changes everything We to bloom this week. I' Lisa.😊
[23:08] And i'm christina raffiti we're tracking breaking news today from wall street to washington this weekend a little bloom.
[23:08] 🎼Your weekend routine what listen stream bringing you up to the minute geopolitical news whenever and more.
[23:08] happens I'm and H in on Turkey. this is bloomloberg.
[23:09] Baalara Atomics, the company building America's first giga sites, secured a billion dollars in its year.
[23:09] LED by Sequoia. The company also closed a $200 million credit facility to build the worlds first vertically integrated.
[23:09] Induial nuclear power Company joining us now is vor Atomic CEO and founder, Isaiah Taylor. You're fresh.
[23:09] Of what now seems with hindsight, a pretty key demo that was a bit of an unlock for this round want.
[23:09] So let's start there. What do you need the funding for and how you' going to deploy it Yeah and thanks for having me on again just a month ago.
[23:09] On the show talking about how we became the first startup in history to make nuclear power, we did that by powering and Nvidia Blackwell ship.
[23:10] It was an amazing event it was the first time that an advanced nuclear reactor has powered AI infrastructure directly, so you're right, it's been an amazing.
[23:10] A months, we made it a lot of progress. and really, what this is about iss giving us the firepower that we need to go build the giga sites.
[23:10] tra nuclear power, this is the reactor standing behind me here, and we have to build a lot more of these in order to make pace with AI A.
[23:10] Building very quickly, we need a lot of power in every direction, and this is really about scale for us, so this is the firepower that we need to go and build.
[23:10] Sequoia led round but Seanun McGui is also going to join your board Why is that important to you, what is it that Sean can give?
[23:10] You other than just the check You know, it's been amazing to spend some time with Sean over the last few months and get to know the rest of the Sequoia team.
[23:10] Partnership, I think Sean's a really really thoughtful person, he spends a lot of time with SpaceX and understanding that business model, and I think one of the lessons to.
[23:10] From SpaceX is really just the power of scale and pace of production, and that's something that we want to do in nuclear as well, so I think that his lessons that he's learned there,
[23:11] Are going to be applied very well to this business. And I'm really excited to walk them into the board. You know, the numbers here are really big, right So it's a billion dollar round.😊
[23:11] 6 billion post money valuation. But with respect, you, this is still early. This is nascent. It is high risk.
[23:11] Technical, how contingent was that demo that you did just on the show a month ago in literally making.
[23:11] round materialize, how much did investors want to see that this worked You're right, This is super early and you this.
[23:11] very long journey for us. we see this as a multi trillion dollar opportunity. we're going to go and build the biggest energy company on earth, So these are the very early days.
[23:11] SureYou know, we have a lot to prove still, we want to prove pace of production, we want to go from building one reactor in a year to six months to.
[23:11] once a month and then even faster than that so you're right, there's still a lot to prove, you know I think that this demonstration was really important we've said for a long time,
[23:12] Baallor Atomics is a hardware company, We're focused on hardware execution and on full vertical integration. And that's really what we've proved here. proved that we can take.
[23:12] we can build a reactor on it, turn it on and operate it, and it's just one small step, but it's the first of many to go and build a nuclear power company.
[23:12] Iaiah, one of the things you and I have not yet discussed is who the customer ends up being you know, you want to build a new.
[23:12] You a power company of the future. you've explained the scale of the tech. You've powered one black world chip. And I think that's important context with the audience.
[23:12] Who do you ultimately sell it to It's a really good question. I think that nuclear and energy in general is a bit noisy right now. You see a lot of deals getting.
[23:12] That maybe our binding, maybe are not, we've been really, really focused and disciplined on this point that we believe this is a hardware execution problem.
[23:12] one needs energy, they need it at any price, they can get it as fast as they can get it, and so really, really focused on just executing in the hardware and bringing new power.
[23:12] OnAnd I think once we get that new power online, that'll open us up to a bunch of really interesting customers who have large balance sheets.
[23:13] Pd on energy in order to build what they see is the future as well. And so there's a handful of customers there that are interested in what we're doing.
[23:13] We're going to stay focused and we're going to stay focused on hardware execution I say I don't know how much you want to talk about finance, I don't know if you're the finance guy.
[23:13] In truth at Valour, but an interesting credit facility, Pamalaki's Aor and JP Morgan.
[23:13] You know it's $200 million facility, but why was that important you really sort of put it high up in the announcement today know, it is important to us.
[23:13] You know, the long term of this business, right, we don't want to be relying entirely on equity and we think that power is a uniquely debt financialable thing.
[23:13] And so moving this company from something that's equity finance to debt finance is a long term goal of ours, and it's really great to be working with partners like A Boorhood.
[23:13] This round and JP Morgan, Presing Cove and Hercules as well coming into it, and we're really starting to demonstrate to the market that this is a business that,
[23:14] Take large amounts of credit, deploy it very well, get the yield on that, and this is how we're going to run the business going forward is that these reactors will be fine.
[23:14] They'll be financed by customers by different ways of putting these tranches together, and that's really how we're going to go and build all of this capex.
[23:14] We're excited to start to demonstrate this with Air. Isaiah, let's end the conversation of talking about America's attitude.
[23:14] Progress on your industry is America moving fast enough so that you can do the things that you want to do on the timelines.🤢
[23:14] That you've set out, look it never feels like we're moving fast enough, right, we want to move so much faster than we are today, China has about 14 nuclear reactors.
[23:14] Under construction right now, and we're just starting to break ground on our next one here. So we need to move faster. if we want to win on AI, we have to go.
[23:14] Faster in all forms of power generation, but especially nuclear, I think nuclear is uniquely well suited to win the AI race. It's clean, it's scale.
[23:14] Very well. You can build lots of it in one place next to a large data center. There's a lot of things that work really well for this, but frankly we do need to move faster and that's why we're so,
[23:15] Focus on the hardware execution Valla Atomic CEO and founder Isaiah Taylor, back on the show, thank you very much indeed.
[23:15] Let's get to the earnings story, stay with AI up next, palantir, set to report its second quarter earnings today after the closing bell, joining us, Bloombergs.😊
[23:15] Iette Chapman leads our coverage of palantier. You know, It's so interesting. We go with these periods of times where we talk about palantier constantly, then maybe it.
[23:15] A little quiet.Ultimately, the company has a lot of momentum, right, What are youbracing for, right It's a pressure cooker.
[23:15] They have a history of always beating or exceeding analyst revenue.
[23:15] Analyst estimates for every quarter on revenue, et cetera, so this is already the expectation that they're going to beat or exceed and so even if they raise their guy.
[23:15] It's kind of like, so what you do this every quarter. And so the big thing that people are going to be watching right now is the concentration of.
[23:16] Sales whether it's in the U, whether it's international this quarter they are expecting really blocked.
[23:16] Busster growth for u commercial sales this has been its strongest area of growth and next quarter for the first time ever.
[23:16] In its history, analysts expect them to have commercial growth, exceed government growth.
[23:16] What they're going to be looking for this quarter is has it continued that trajectory and how are they doing internationally, which is something that.
[23:16] We've talked about Markon and some other people on, so that's something that they're going to be looking at it's a really big point because you know in the news headlines, palentic.
[23:16] Comes up in the context of its work with the Pentagon or with different arms of the U government in a defense context. But they also have a lot.
[23:16] Commercial case studies, so that's the moment in time it flips to commercial, but give us the basics of like what palents business does, right we always have to remind us.😊
[23:16] what isin, it is AI software that allows people to make the best use of the.
[23:17] Information that they have to gather data from all across different sources that they have to integrate it.
[23:17] Make sense of it, map it, and then throw an LLM up against it, whether it is you know from from.
[23:17] whichever company whether they exactly, and so they've worked a lot on showing different partnerships that they have.
[23:17] With some of the major players, that's something that they're going to talk about today, i'm sure, along with partnerships with NviD and some of the rest.
[23:17] They have gotten a lot of attention for their work with defense, but it's actually the commercial.
[23:17] Which is their fastest growing and so we'll see how they do Alex carp, the CEO is.
[23:17] y he also talks about the markets sometimes, but is there a clear palantist stock story of this year of the last year, like how was the market?
[23:18] Right the market had massive expectations at the end of last year the stock was up.
[23:18] Almost up to 200 or so and since then it's plummeted from that high, it was trading at something like 247 x earnings then.
[23:18] Now it's down to about 67 x earnings, which is down from its five year average of 98 x earnings, but it's still one of the.
[23:18] Most expensive stocks on the S&P. So a lot of investors are wondering where that disconnect is and whether it's.
[23:18] Really worth what it said it was going to be able to deliver this seems like a big one, a big one for markets bloombergette chapman who covers Panant but also.
[23:18] 🎼defense tech, venture capital, thank you very much indeed now coming up Apple's changing the way you buy its devices, why monthly payment?😊
[23:18] 🎼This could become the new normal power on, this is Bloomberg Tech.
[23:19] 🎼Bringing you up to the minute news whenever and wherever it happens. I'm Haslenda I in in Mumbai this.
[23:19] At a time when policy in washington is driving wall street, we draw a distinction between the rhetoric.
[23:20] And the action bringing you market moving guests and original reporting, this is Bloomberg surveillanceance.
[23:20] Mind being called an ideological lunatic or a bunch of left wing nut jobss, you know, I've been called worse things than that.
[23:20] 23 and me, I think is so incredibly valuable, we're coming back.
[23:20] What magazines were you reading as a kid I don't think I was reading magazines until I was in them so is it paradise ahead or Gattica?
[23:20] Maybe somewhere in the middle.
[23:20] 🎼World of politics to the world of business. Joe Matthew and Ka lines deliver news insight and analysis live two time.
[23:20] Today day, every trading day, balance of power at one and five Eastern only on bloomomberg.
[23:21] Pret check on Amazon shares at all time highs, they've joined the $3 trillion market cap Club.
[23:21] 🎼And this is the biggest two day game they've posted since 2009 after earnings, later in the hour, we're going to speak to AWSCOO Matt Garman, AWS.😊
[23:21] 🎼Amazon's key cloud computing unit, Apple is making a big bet that customers would rather pay by the month than all at.
[23:21] It's the new Apple upgrade program designed to make you upgrading your device feel more like a subscription.😊
[23:21] Power on from this weekend. It's been really interesting to see the rollout of this. What are you trying to capture in this power on Where do you think.😊
[23:21] This is headed Yeah, the fact of the matter is, is that you're going to see a change over time from sticker prices to subscription price.
[23:21] The carriers have been trying this for years. It's been extremely successful for the carriers in the US internationally. Obviously Apple.😊
[23:22] Gotten on a really good schedule with its new products, obviously new iPhones, and Apple watches every September, the iPad ca.
[23:22] This has really become fall and spring, the Mac, depending on which Mac you get, you're thinking fall or spring as well.
[23:22] There are upgrade cycles for consumers, right maybe I'm going to get a new phone every three years, maybe I'm going to get a new computer every five years, but making it a subscription.
[23:22] Is going to get people into getting a new phone more frequently. They could give those phones then back to Apple. Apple could resell them.
[23:22] Secondary markets that can reuse the parts for new devices,, which creates a big cost savings and improve margins for them there too as well.
[23:22] seary phone business, and then the consumer is paying on a recurring schedule, so the payments to Apple never stop and over the long term.
[23:22] ing to see Apple creating far more revenue in profit because of this program. In fact, Wall Street has been begging Apple to do this for over half a decade.
[23:22] So it's a long time coming So this weekend I did my first Apple upgrade, I took my 2020 Mac Air with an M1 chip.
[23:23] I took the latest Macair 13 inches with an M5 chip and I did the math and it's a great deal, basically I can trade in still.😊
[23:23] And still do the Kna Apple upgrade, the thing that jumped out is I've got to wait three weeks for that Mac Air, that's still an issue.😊
[23:23] Yes, so as we wrote in Power on as well this weekend, the Mac Studio and the Mac Min, those are no longer the only max with extreme supply.
[23:23] Now it's the most popular Mac, which is of course, the MacBook Air and new orders of the MacBook Air are not arriving at this point until early September.
[23:23] So you're seeing the supply chain shortage shift from niche computers to the company's most popular computer, so that's not necessarily great, and so what they're doing.
[23:23] In terms of their marketing online and in stores, they're pushing people towards the base level M5 MacBook.
[23:23] Pro, the 14"ch, which they have a lot of supply of still, and they want to sort of dwindle that down because they're launching the new M6 version.
[23:24] Just a couple months from now Mark just very quickly, we have a story on the terminal about how Apple's Siri AI stacks up against the new Cha UBT.
[23:24] What do we need tonight Apple has a huge advantage by having Siri AI built into every iPhone now that can run.
[23:24] I 27 and the latest large language models, it's competent, right, the promise of Siri has been there for 15 years, it's never really worked.
[23:24] Now it works, it's nowhere near as good or as capable as ChayBT, but the floor has improved tremendously.
[23:24] 🎼Floombo Smart Goman with the latest power on thank you very much now coming up AWSCO Matt Gom joins us to talk about the Business blockbuster Corder.😊
[23:24] 🎼Amazon hits the $3 trillion market value Club this is what your markets look like it's hard time here in San Francisco, this is Bloomberg Tech.😊
[23:25] Are you.
[23:26] different president from your father I have to be it's a different time. There's fundamental lessons that I still hold.
[23:26] Close to my heart, but in terms of the practical things that you do, very few of those things apply any longer.
[23:26] How' mom Oh she she's in her 90s Yeah, but I always say she's 95 years old, she didn't get the memo, she doesn.
[23:26] She doesn't believe she's 95 years and she's still very active, I mean I still get calls from her and she says.
[23:26] You need to do this, you need to do that and she doesn't let her age hold her back try to advise you.
[23:26] As president of the Philipes constantly constantly I always learned something new with my mom in my family, she's the realpoli.
[23:26] 🎼And she said, naturally, my father was the intellect. He was a statesman, and we always.
[23:26] So that he lived in his head, you know, it was very, very, very deep thinker that's what I learned my mother on the other hand.
[23:27] She has this uncanny ability to connect with any kind of person, any person she runs into industry the.
[23:27] King's presidents, anyone, her childhood friend, she doesn't know how to fight, she's very bad at conflict.
[23:27] 🎼Welcome to the world of decentralized.
[23:27] Finance Bloomberg is covering all things crypto, the people, the transactions, and the technology.
[23:27] Bloomberg crypto Tuesdays only on bloomberg me each week on Wall Street.
[23:27] Capitalism from business, markets, economics, tech and climate, more than what you need to know, it's what you need to think about.
[23:27] Bucke back to Bloomberg Tech, Amazon has blown past $3 trillion in market value after its cloud unit.

### [E1804] Bloomberg | 2026-08-03 19:20 UTC | 盘中
OpenAI, Anthropic, Google to Join White House AI Safety Meeting
URL: https://www.bloomberg.com/news/articles/2026-08-03/openai-anthropic-google-to-join-white-house-ai-safety-meeting
The Trump administration plans to host artificial intelligence companies at the White House on Tuesday to discuss a new US framework for conducting voluntary safety tests of AI models, according to people familiar with the matter.

OpenAI, Anthropic PBC and Alphabet Inc.’s Google are among the AI developers planning to attend, according to the people, who described the meeting on condition of anonymity because it hasn’t yet been announced. The recently completed framework springs from a June executive order by President Donald Trump on AI cybersecurity that outlined an opt-in approach to safety reviews of models and greater efforts to shore up critical computer systems.

The framework has not yet been released publicly, and Trump’s order indicated that some of its benchmarks would remain confidential. Nonetheless, it marks the Trump administration’s latest step toward addressing AI security issues, an initiative that accelerated in April after Anthropic warned that its Mythos model was adept at finding computer vulnerabilities and strictly limited its release.

Spokespeople for the White House and the three companies declined to comment. The Information reported earlier on the Tuesday’s meeting plan.

Read More: Anthropic, OpenAI Cyber Failures Point to US Security Risks

Adding urgency to the effort, OpenAI and Anthropic have disclosed in recent weeks that a handful of their models escaped secure testing environments and hacked third-party organizations.

At the same time, both companies have chafed at what they see as an inconsistent approach to AI safety reviews, and are looking to the new framework to bring more consistency to what had been initially presented as a voluntary program.

In June, weeks after Trump signed the order and signaled a largely hands-off approach to testing, the Commerce Department barred Anthropic from sharing its two most powerful models with foreign nationals, citing concerns that safety guardrails could be circumvented. The company temporarily disabled access to its Fable 5 model and its selectively shared Mythos 5 model, eventually making changes to address the administration’s objections.

Read More: US Considers Creating Finra-Like Watchdog to Vet Top AI Models

Partly in response to the industry’s complaints of an ad-hoc approach to model reviews, Treasury Secretary Scott Bessent proposed creating an independent regulatory agency that would give companies a significant say over safety reviews. The contours of that plan align with a proposal from Google DeepMind’s leader Demis Hassabis, who was in Washington last month to meet with policymakers.

Bessent’s model would set up a system similar to the Financial Industry Regulatory Authority, but apply to the artificial intelligence industry, Bloomberg News has reported. It remains unclear if that plan would be part of the AI cybersecurity order implementation or a separate effort.

Tuesday’s discussions will take place less than two months before Trump meets with his Chinese counterpart Xi Jinping for a high-profile summit in Washington where the competition for AI leadership between the world’s two largest economies will take center stage.

Read More: Alibaba Adds to China AI Breakthroughs With New Qwen Model

US officials are now grappling with how to respond to a wave of new AI models from Chinese firms that rival top offerings from American labs. Last month’s release of the new Kimi K3 model by Chinese startup Moonshot AI undermined investor confidence in the durability of the US lead in AI while raising questions about the hundreds of billions of dollars in data center investments. Kimi, which costs a fraction of the price of top US offerings, nears their performance in some domains.

That challenge was highlighted by recent moves by other Chinese developers. Over the weekend, DeepSeek expanded access to its latest model and on Monday, Alibaba Group Holding Ltd. unveiled its own new system, Qwen3.8-Max, which the company said is on par with Anthropic.

The slew of releases has fomented tech industry debate over how the US should respond to open-weight AI models from China that cost far less than the best US products while approaching their performance in some domains.

Silicon Valley leaders including Nvidia Corp. Chief Executive Officer Jensen Huang have insisted that open-weight models, which allow users to download and customize the technology, are good for the long-term development of AI and help bolster its security. In public statements, with dozens of companies signing in support, they’ve urged the US to refrain from restricting open models.

Read More: Moonshot’s Kimi Uses 20,000 Nvidia Chip Cluster From Alibaba

Other key tech industry figures, including Anthropic CEO Dario Amodei, have suggested that the gains made by Chinese AI models violate US rules, concerns echoed by Bessent and other top US officials.

Last month, White House Director of Science and Technology Policy Michael Kratsios accused Moonshot of illegally acquiring Nvidia’s advanced Blackwell chips, the most powerful AI accelerators on the market. It’s unclear how many of these processors the US government believes Moonshot has. Bloomberg has reported that Moonshot has a computing power agreement with Alibaba for the use of about 20,000 Nvidia chips, forming a substantial chunk of the overall capacity Moonshot uses for its Kimi models.

Kratsios also said Moonshot improperly extracted data from US models through a process known as distillation, a style of reverse engineering that US AI companies have warned officials about for most of this year. Using this technique, researchers feed the outputs of one model into a new version, a process that turns out to be an efficient way to recreate some of the parent model’s capabilities.

### [E1853] SemiAnalysis | 2026-08-03 19:42 UTC | 盘中
Kimi K3 took the world by storm at its announcement, sweeping leaderboards and establishing itself as the open frontier model. While the community is eager to understand how Kimi K3 works, many have been surprised by the unconventional tech
URL: https://newsletter.semianalysis.com/p/kimi-k3-the-manos-the-mythos-the
Kimi K3 took the world by storm at its announcement, sweeping leaderboards and establishing itself as the open frontier model. While the community is eager to understand how Kimi K3 works, many have been surprised by the unconventional techniques driving its performance. This article serves as a primer to understanding the core techniques of the Kimi K3 model architecture.

# Kimi Delta Attention

Kimi Delta Attention (KDA) is the linear attention layer in Kimi K3’s hybrid attention mechanism. We trace the origins of KDA, starting from linear attention, DeltaNet, Gated DeltaNet (GDN), then to KDA.

## Linear Attention

The derivation of linear attention stems from **removing the softmax operation in the standard softmax attention**. Below we compare the iterative inference formulas, which show the computation of the output vector at token position t:

By removing the softmax operation, we can reorder the operations and reduce the computation complexity of attention from quadratic to linear:

The new equations are as follows:

Vectors q, k, v, have dimensions *L* by *d*. The computational complexity of both equations are O(*Ld*²), thereby making the computation linear. Comparing the new equations with softmax attention’s equation, we see that **softmax attention requires accessing all past key and value vectors**, whereas **linear attention compresses all past key and value vectors** **into one hidden state S**.

We reinterpret the new equations as an **online learning objective**. We view matrix S as an associative memory that stores the associations between key vector k and value vector v, and we retrieve v by multiplying S with k. We can then interpret the first equation as **continuously updating the matrix S at every position to perfect the retrieval**. Finally, we can interpret the vt @ kt.T term as the gradient of loss function -(S @ kt.T) @ vt with respect to S.

## DeltaNet

Under the online learning objective view, we see the values of matrix S will grow unboundedly: old and new information gets blurred together in S as the sequence grows, which destabilizes learning. Without softmax giving well-scaled and bounded outputs, linear attention typically lags behind softmax attention on long-range recall tasks.

DeltaNet improves upon linear attention by **changing the loss function to minimizing the L2 norm of the value retrieval**. Unlike linear attention’s loss function, DeltaNet’s loss function regularizes the growth of S. This creates a new matrix S update rule, the Delta Rule, as below:

Source: [Linear Attention and Beyond (Interactive Tutorial with Songlin Yang)](https://www.youtube.com/watch?v=d0HJvGSWw8A)

The Delta Rule becomes the basis of DeltaNet’s attention equation:

Conceptually, Sₜ-1 @ kₜ - vₜ represents the associations irrelevant to the current key and value, and DeltaNet performs targeted removal of those associations.

## Gated DeltaNet

GDN and KDA are adaptations of DeltaNet. Gated DeltaNet applies the LSTM forget gate *alpha *on the matrix S, allowing the model to control memory lifespan with weight decay. KDA further expands *alpha* into a diagonal matrix that enables fine-grained per-channel memory decay and positional awareness.

Thanks for reading SemiAnalysis! This post is public so feel free to share it.

[Share](https://newsletter.semianalysis.com/p/kimi-k3-the-manos-the-mythos-the?utm_source=substack&utm_medium=email&utm_content=share&action=share&token=eyJ1c2VyX2lkIjo1MTA4NTkyNDAsInBvc3RfaWQiOjIwOTE4OTA1NSwiaWF0IjoxNzg1Nzg2NDQ5LCJleHAiOjE3ODgzNzg0NDksImlzcyI6InB1Yi02MzQ5NDkyIiwic3ViIjoicG9zdC1yZWFjdGlvbiJ9.ALObgJBeAoF2IRxAF-bLK7AUnDp-zUyH94vfBDGmicQ)

## FlashKDA Algorithm

Moonshot developed FlashKDA, their custom kernels for KDA, and [open-sourced it](https://github.com/MoonshotAI/FlashKDA/tree/master). Here we explain the algorithm and derive the arithmetic intensity.

### Algorithm

First, let’s start from an alternative formulation of the recurrence formula:

`u_t = beta_t * (v_t - (D_t @ S_t-1).T @ k_t)
S_t = D_t @ S_t-1 + k_t @ u_t.T
o_t.T = q_t.T @ S_t`
```

Here, D_t is the diagonal matrix of the alpha forget gate, and u_t is the delta in the delta rule. For decode, the kernel roughly follows the formula. For prefill, we parallelize the operation by unrolling the recurrence formula in chunks of tokens, in order to efficiently execute the operations on GPUs. Assume we unroll token i to j, and the starting state is S_i-1, we get:

`S_j = D_j:i @ S_i-1 + sum(D_j:t+1 @ k_t @ u_t.T, t=i:j)
o_j.T = q_j.T @ S_j
= q_j.T @ D_j:i @ S_i-1 + sum(q_j.T @ D_j:t+1 @ k_t @ u_t.T, t=i:j)`
```

D_j:i refers to the cumulative decay from token i to j: D_j @ D_j-1 @ D_j-2 @ … @ D_i. In FlashKDA’s matrix form, the formula becomes:

`S_out = D_j:i @ S_in + K_restore.T @ U
M_qk = tril(Q_decay @ K_inv.T)
O = Q_decay @ S_in + M_qk @ U`
```

The vector to matrix mapping is as follows:

-

S_in refers to the state at the starting position of a chunk

-

S_out refers to the state at the end position of a chunk

-

K_restore is the matrix form of D_j:t+1 @ k_t

-

Q_decay is the matrix form of q_j.T @ D_j:i

-

Q_decay @ K_inv.T the matrix form of q_j.T @ D_j:t+1 @ k_t, derived from (q_j.T @ D_j:i) @ (D_t:i^-1 @ k_t)

-

M_qk is the causal mask, so it’s a lower triangular matrix

U is the matrix form of unrolled u_t. To compute this, we apply UT transform and compute the following:

`B = Diag(beta) @ (V - K_decay @ S_in)
L = StrictTril(Diag(beta) @ K_decay @ K_inv.T)
U = (I + L)^-1 @ B`
```

Please consult [Songlin Yang’s blog post](https://sustcsonglin.github.io/blog/2024/deltanet-2/) and [Kimi Linear paper section 3.1](https://arxiv.org/abs/2510.26692) for the full derivation. Note that here U corresponds to the pseudo-value term in the Kimi Linear paper.

Implementation-wise, FlashKDA launches 2 kernels: K1 and K2. K1 prepares chunk-level tensors in parallel, including:

`a = exp2(cumsum(g))
K_decay = Diag(a) @ K
Q_decay = Diag(a) @ Q
K_inv = Diag(a)^-1 @ K
K_restore = a[-1] * K_inv
L = StrictTril(Diag(beta) @ K_decay @ K_inv.T); INV = (I + L)^-1
M_qk = tril(Q_decay @ K_inv.T)`
```

Here `a` is the cumulative decay, where each element is the cumulative decay at a token position.

K2 performs chunk-level recurrent computation:

`U = INV @ Diag(beta) @ (V - K_decay @ S)
O = Q_decay @ S + M_qk @ U
S = Diag(a[-1]) @ S + K_restore.T @ U`
```

### Complexity Analysis

Here we analyze the complexity of an attention head. For decode, the critical path computations are:

-

D_t @ S_t-1: Element-wise multiplication, D × D

-

S_t-1.T @ k_t: D × D × 1

-

k_t @ u_t.T: D × 1 × D

-

q_t.T @ S_t: 1 D × D

The decode kernel roughly performs 7*D² FLOPs.

Reading and writing the FP32 recurrent state dominates the memory traffic, so the memory traffic is roughly 8*D² bytes.

For prefill, the critical path of K1 is at computing L, INV, and M_qk.

-

L: C × D × C

-

INV: [Neumann factorization](https://github.com/MoonshotAI/FlashKDA/blob/1ce47ea3bb22c84eb9cc665028399cf35e8ffb0b/csrc/smxx/utils.cuh#L190), performs 6 C × C × C matrix multiplications

-

M_qk: C × D × C

For K2,

-

K_decay @ S: C × D × D

-

Q_decay @ S: C × D × D

-

M_qk @ U: C × C × D

-

INV @ B: C × C × D

-

K_restore.T @ U: D × C × D

Combining K1 and K2, FlashKDA performs 12*C^3 + 8*C²*D + 6*C*D² FLOPs. Since we analyzed at the chunk level (chunk size C), assuming sequence length T >> C, the overall FLOPs is T/C * O(C*D²) = O(T*D²).

For memory traffic:

-

K1 read Q, K, g: C × D

-

K1 write and K2 read Q_decay, K_decay, K_restore: C × D

-

K1 write and K2 read INV, M_qk: C × C

-

K2 read V and write O: C × D

-

K2 read and write S once per kernel: D × D

In total, FlashKDA accesses 3 * 2*C*D + 2 * (3 * 2*C*D + 2 * 2*C*C) + 2 * 2*C*D = 8*C² + 22*C*D bytes. At the kernel level, it accesses T/C * (8*C² + 22*C*D) + 8*D² ~ O(TC + TD + D²).

This concretely shows that the computational complexity of KDA:

-

Prefill: Linear to sequence length for both computation and memory

-

Decode: Constant to sequence length for both computation and memory

# Kimi Linear

Moonshot trained Kimi Linear models as proof of concept for their KDA design, so we can infer Kimi K3’s architecture design from Kimi Linear. Comparing the K3 release tech blog with Kimi Linear, we see Kimi K3 shares the shared expert count, the hybrid linear attention ratio, and the general attention module design.


_Source: SemiAnalysis_


_Source: [Kimi K3 Tech Blog](https://www.kimi.com/blog/kimi-k3)_

The diagram above shows the operations performed on the inputs of KDA. For the query, key, and value, we apply linear transformation and short convolution. Applying short convolution effectively capturing local token dependencies, and doing a left padding convolution avoids breaking causality. We additionally apply L2 norm to the query and key to stabilize the eigenvector of the transition and the output matrices. For the decay memory gates, alpha is a low rank projection, and beta is a down projection. The KDA output is normalized per head and controlled by an output forget gate, implemented as a linear transformation in K3, instead of a low rank projection in Kimi Linear. Finally, we apply a linear layer to mix per-head information.

Kimi Linear interleaves KDA with full attention Multi-head Latent Attention (MLA). Kimi Linear showed that 3:1 is the ideal KDA to MLA ratio that balances performance and efficiency. KDA also serves as a strong position-aware operator, replacing the RoPE in MLA.

Keeping MLA as full attention is an interesting choice, as all other open weight models move to Grouped Query Attention (GQA). MLA uses an absorption trick to reduce the computation of a decode step at the cost of extra computation during the prefill step. This is a sensible trade-off for decode-dominant reasoning workloads, but for prefill-dominant agentic workloads, extra computation becomes a high cost with little benefits. As a result, all frontier open weight models use GQA-based attention mechanisms: GLM 5.2 DeepSeek Sparse Attention, DeepSeek V4 Compressed Sparse Attention, MiniMax M3 MiniMax Sparse Attention, and MiMo V3 HySparse are all based on GQA. We suspect Moonshot’s future models such as Kimi K4 will feature attention mechanisms that replace MLA.

SemiAnalysis is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
**

# KV Cache Efficiency

We argue that **one should not infer KV cache efficiency solely based on KV cache space complexity**. KV cache size is not a standalone factor but a property of the model design: no open weight models are released with static KV cache compression techniques, and model architecture inference efficiency affects KV cache efficiency. The effects of KV cache size also vary, depending on the total memory capacity of a deployed model instance. For example, deploying a model with wide expert parallelism has very different memory profiles than doing so with tensor parallelism, which affects the memory capacity left for KV cache. Thus, we propose considering both the model architecture system efficiency and the KV cache size to understand the KV cache efficiency, and we quantify that with **KV throughput**.

## KV Throughput

KV throughput is defined as KV cache size divided by the prefill time (Time to first token), given a specific sequence length. KV throughput represents the minimum bandwidth required to reliably serve a model with PD disaggregation, but it is also a good proxy for understanding KV cache efficiency. Prefill time encapsulates the efficiency of the model architecture, and as the sequence length increases, we will see the memory-bounded and the compute-bounded situations. As shown in the table below, we can see the benefits of hybrid linear attention become more pronounced as sequence length increases.


_Source: [Prefill-as-a-Service: KVCache of Next-Generation Models Could Go Cross-Datacenter](https://arxiv.org/abs/2604.15039v1)_

This is also a good way to understand the bandwidth requirements of KV cache offloading to different memory tiers in a cluster.

## KV Cache Residency

The location where KV cache is stored follows the memory hierarchy. First, KV cache resides in HBM, the fastest memory in a GPU cluster, consuming whatever capacity is left by model weights and activations. As KV cache size exceeds the HBM capacity, it spills into server DRAM, a higher capacity but lower bandwidth memory pool. Finally, when KV cache exceeds DRAM capacity, it spills to disk storage such as SSD. This is analogous to the computer architecture cache hierarchy: register, cache memory, main memory, disk storage.

The analogy continues for memory coherency. Popular distributed KV cache framework Mooncake Store supports write-through and write-back policies for KV cache loading. Mooncake Store features a distributed KV cache pool that makes all KV cache visible to all workers. Implementing write-through policy between DRAM and the lower-level distributed KV cache pool offers multiple benefits in multi-node scenarios, including sharing prefix cache across nodes, avoid KV cache duplication for tensor parallelized MLA, and KV cache redundancy when a node goes down.


_Source: SemiAnalysis_

## KDA Prefix Cache Management

At each token position in a request, Kimi K3 KDA’s recurrent state is fixed in size, whereas standard attention KV cache grows with sequence length. This KV cache space reduction comes at the cost of complicating prefix caching, especially when Kimi K3 is a hybrid attention of KDA and MLA.

Roughly speaking, modern inference engines identify prefix cache hits by matching the longest token prefix in the existing cache.


_Source: SemiAnalysis_

Identifying the longest prefix becomes a problem for linear attentions like KDA. Without prior knowledge of where the boundary of a prefix is, we will have to cache KDA’s recurrent state at every token position. This means every token has a cache, and the KV cache memory usage regresses to growing with sequence length, defeating the purpose of using linear attention. To tackle this problem, Moonshot saves recurrent states at a coarse granularity, e.g. vLLM caches every 32K tokens. vLLM additionally caches at prompt boundaries, since for agentic workloads, a new turn typically starts at the end of a prompt.


_Source: [Kimi K3 Is Here: Efficient Day-0 Support on vLLM](https://vllm.ai/blog/2026-07-27-k3)_

This shows that even though linear attentions like KDA greatly reduce KV cache memory consumption, **realistically during serving, they do not consume a constant amount of KV cache memory**.

# Attention Residuals

## Residual Connections

Residual connections are one of the key innovations that allowed us to build bigger deep neural networks through scaling model depth. The deeper the neural network, the more expressive they become but training them naively is hard. Signals from the earlier layer need to be preserved till the last layer and gradient need to survive from output to first without vanishing.

Instead of modeling whole networks as a single function, passing information only through nonlinear transformations, residual networks connect smaller blocks with identity paths. Each block fᵢ ​ learns a change to its input xᵢ, given by the recurrence:

The identity mapping allows features to carry from shallower units to any deeper unit and gives the gradient a path highway so they do not vanish.

While residual connections allow us to build deeper networks, they come with challenges.

Early layers heavily influence residual stream to have effect on final output. Because of which residual stream has irreversible information loss with increasing depth. Later layers increase output gain to have effect on this modified residual stream which can destabilize training. Another variant like highway networks allow gating mechanisms for information flow but they suffer from the same crucial problem. Layers don’t have selective access to information from earlier layers.

## Recurrence In Time and Depth

Sequence modeling dominated by recurrent neural networks had the same recurrence formulation.

Where each step has identity mapping with previous state for direct information flow and the sequence model faced the same challenge: depth in the time axis dilutes signal.


_Source: SemiAnalysis_

Attention machines transformer removed this constrained by retrieving any token in past with powerful and expensive attention mechanism

## Attention on residual stream

Motivated by attention mechanism in sequence modeling, kimi developed attention residual, where they take attention over depth blocks,


_Source: [Attention Residuals](https://arxiv.org/pdf/2603.15031)_

Standard causal self-attention computes the output of token $t$ as a weighted sum of previous token representations:

Attention Residuals use the same attention mechanism, but replace the sequence dimension with the depth dimension. Instead of attending over previous tokens, each layer attends over representations produced by previous layers.


_Source: SemiAnalysis_

Unlike standard attention, the query is a learned parameter for each layer rather than being generated from the current token.

For each layer ℓ, we define:

Each colored block represents the output token representation from a previous transformer layer. Just as standard causal self-attention performs softmax attention over tokens in the sequence, Attention Residuals perform softmax attention over the representations produced by previous layers.


_Source: SemiAnalysis_

Attention residual allows the model to get fine grained control over what inputs to pick from past layers making the model more expressive.

## Block Attention Residuals

Attention residual need to all past layer outputs for attention. For large models distributed over many GPUs this creates O(Ld) communication overhead. To overcome this block attention residual dividends layers L into N blocks of S layers. Block AttnRes applies attention over completed block outputs and for current block its evolving partial sum.


_source: SemiAnalysis_

Block Attention has minimal tread over full attention residual but they cut down communication from O*(Ld)* to O*(Nd).*

Let bₙⁱ denote the partial sum over the first *i* layers in block *n*, such that

For the *i*-th layer in block $n$, the available block representations are
1.\n\\end{cases}\n","id":"FAVEMAHKYQ"}" data-component-name="LatexBlockToDOM">

Unlike standard attention, the query is not input-dependent. Each layer learns a query vector:

Attention weights over the available block representations are computed as

The output is the weighted sum of previous layer representations

Rather than depending only on the residual stream to preserve information, Attention Residuals give every layer direct, selective access to earlier representations. This block based variant of attention residuals greatly reduces communication overhead while having competitive performance.

Block residuals show better scaling compared standard residual connection achieving 1.25× compute efficiency. Consistently lower validation loss compared to baseline and gap widening with decay phase. Unlike standard residual networks where output magnitude increases as depth increases. selective aggregation of block attention has bounded output. And consistent gradient magnitude.

[Subscribe now](https://newsletter.semianalysis.com/subscribe?)

## Training

Unlike standard residual networks, attention residuals need all N-1 block input for computation of the Nth layer. This becomes a problem for pipeline parallelism as all N layer blocks output need to be transferred across stages.

With clever cross stage caching and activation checkpointing, Kimi reduced overhead to only 4% compared to standard architecture for pipeline parallelism.

### Cross-stage caching

For *P* physical stages and *V* virtual stages. Each block *N* needs *C=PV* communication for each chunk. Naively this needs transferring all accumulated blocks for each stage. This is quadratic cost growth for each physical and virtual stage

This high communication can be reduced by caching input across virtual stages. Blocks computed in earlier layer can be stored in local memory,


_Source: SemiAnalysis_

For the first virtual stage all block embedding needs to be transferred in the physical stage, each completed block is stored on respective rank. For all subsequent virtual stages all cached blocks can be reused for computation. Only the block not present on rank need to be transferred for attention to residual computation.


_Source: SemiAnalysis_

These split communication costs for first and subsequent virtual stages. For the first virtual stage its need incur the same quadratic cost for all physical layers. In subsequent virtual stages we need cached inputs from local devices and Transfer of only PNp chunks needed cutting down total communication from O*(C)* to O*(P)*

The cutdown of communication is directly proportional to virtual stages V. Because of this for full stage of one forward and backward pass all computation and communication can be overlapped

### Memory overhead

Due to cross stage caching all blocks are stored once across all V virtual stages. With Activation checkpointing all inter-block chunks for attention are eliminated. Each stage activation checkpoint *Pl* matches memory size of H*l* of standard architecture and has no extra memory cost.

## Inference

Because Attention Residuals need the output of all previous blocks to compute attention, a naive implementation has excessive memory accesses. To reduce overhead, inference is split into two phases which mirror prefill and decode stages of autoregressive attention. This computation is divided into inter-block attention for completed blocks and intra-block attention for evolving attention in the running block.

#### Phase 1: Parallel Inter-Block Attention


_Source: SemiAnalysis_

During decoding, we have to output the completed block and the query vector learned per layer. All inter block layers simultaneously with a single batched query against the completed block representations, returning both outputs and softmax statistics which can be reused for further computation. This phase is similar to prefill phase decoding

#### Phase 2: Sequential Intra-Block Attention


_Source: SemiAnalysis_

This phase is analogous to the decode phase, Similar to flash attention, evolving sum can be computed with online softmax for intra blocks combined with precomputed inter block results. Which reduces redundant memory access.

With this two phase design, the IO footprint is similar to standard residual architecture, with only the addition of phase ones inter block computation, amortized by batching all queries in the block.

# LatentMoE

LatentMoE compresses the routed tokens before the dispatch operation, and then decompresses them after the aggregation operation. In Kimi K3’s Stable LatentMoE, they apply an RMSNorm before the up-projection (decompressing) operation to reduce sensitivity to scale variations and improve model performance.


_Source: [Kimi K3 Tech Report](/api/semianalysis-db/file/209189055/a3a77ef3a0379cd00336ead3.pdf)_

Here we explain the design principles behind LatentMoE regarding MoE communication. As shown in the LatentMoE paper, the communication volume is proportional to total routed tokens *t*, number of active experts *K*, and expert input dimension *d*, while being inversely proportional to the expert parallel size *E*. This is potentially the reason behind Kimi K3’s latent MoE dimension size and active expert count configuration. Kimi K2 series feature 8 active experts with input dimension size 7168, so Kimi K3’s latent input dimension size being 3584 (half of 7168) would allow the active expert count to double to 16 without increasing the communication volume.

However, the **ratio of communication to computation** **time** is arguably more important for estimating system efficiency (Discussions [here](https://x.com/chhillee/status/2077966168304787769) and [here](https://x.com/chhillee/status/2078130513546723531)). The ratio indicates the roofline of how well MoE kernels can overlap communication with computation at a throughput-bound regime, and **expert intermediate dimension size** is the only model configuration that affects the ratio. Specifically, increasing the expert intermediate dimension size would decrease the ratio, meaning that the theoretical maximum fraction of communication that can be hidden is higher. Here we derive the formula:

-

*t*: total input tokens across the expert parallel (EP) domain

-

*K*: number of active experts per token

-

*N*: number of total experts

-

*E*: Ranks in the EP domain

-

*d*: Expert input dimension

-

*m*: Expert intermediate dimension

-

*P*: Aggregate bytes communicated per activation element (dispatch + combine)

-

*F*: Effective FFN expert (modeled as SwiGLU) computation throughput per GPU, FLOP/s

-

*B*: Effective uni-directional network bandwidth per GPU, B/s

-

Assuming uniform expert routing, each GPU is assigned *t * K / E* tokens

-

Assuming uniform expert routing, an average *1 / E* tokens are local to the source GPUs, so each GPU dispatches *(t*K/E) * (1-1/E)* tokens

-

Each token is a d dimensional vector, so the communication volume per token is *d * P*

-

The communication volume per GPU is *(t*K/E) * (1-1/E) * d * P*

-

The communication time T_comm = *(t * K * d * P) / (E * B) * (1-1/E)*

-

The SwiGLU computation involves 3 matrix multiplications:

-

Up (First) projection: *d* to *m*

-

Gate projection: *d* to *m*

-

Down (Second) projection: *m* to *d*

So the computation is *2*d*m + 2*d*m + 2*m*d = 6*d*m* FLOPs per token

-

The computation time per GPU is *T_comp = (6*d*m) * (t*K/E) / F*

-

The communication to computation time ratio is

*T_comm / T_comp*

*= ((t * K * d * P) / (E * B) * (1-1/E)) / ((6*d*m) * (t*K/E) / F)*

*= (P*F) / (6*m*B) * (1-1/E)*

We believe this formula also motivates an increase in expert intermediate dimension to 3072 in not just Kimi K2 to K3, but all recent open weight models, including DeepSeek V4 Pro, MiniMax M3, MiMo V2.5 Pro, and Inkling. As hardware improves and expert weight precision reduces to save memory capacity, the compute throughput increases, so one way of reducing the ratio is by increasing the expert intermediate dimension.

## Quantile load balancing (QB)


_Source: [Kimi K3 report](/api/semianalysis-db/file/209189055/a3a77ef3a0379cd00336ead3.pdf)_

Many previous load balancing methods require careful hyperparameter tuning. Quantile balancing is hyperparameter free aux-loss free load balancing technique developed by Jianlin Su in [Feb 2026 blog post](https://kexue.fm/archives/11619)

Base principle QB is the same as auxfree load balancing where router biases are updated dynamically based on the system’s load. But instead of updated bias by some small coefficient like aux-free lb, QB directly computes the next bias from the distribution of router scores relative to routing cutoff threshold. Bias updates become small naturally when the router balances load evenly.


_Source: [Kimi K3 report](/api/semianalysis-db/file/209189055/a3a77ef3a0379cd00336ead3.pdf)_

QB tries to find the bias that would have approximately balanced under the current cutoffs and routing on the current batch, solving constraint optimization problems and applying these updates for the next batch. The first constraint is that each token is routed to exactly k experts. The second constraint is a batch of m tokens each picks k experts, gives (mk) assignments in total, to spread load evenly across n experts each expert should process $q=mk/n$ tokens.

Each token finds the cutoff threshold as the (k+1)-th highest biased router score and uses it to calculate the bias update needed to balance load for each expert. For each expert, QB sorts the margins between its router score and every token’s cutoff. It sets negative bias to q+1 the largest margin, leaving exactly q margin above the threshold. Since q/m=k/n, this is (1-k/n) quantile of the margin, which is why it’s called Quantile Balancing.

SemiAnalysis is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
**

# Inference performance

We are actively tracking Kimi K3’s inference performance on [InferenceX](https://inferencex.semianalysis.com/).

As of 30th July, all providers on OpenRouter have a floor of $3 per million tokens input and $15 per million tokens output. Both Nvidia and AMD had Day 0 recipes on vLLM, boasting DRAM offload and DSpark speculative decoding.


_Source: [OpenRouter](https://openrouter.ai/moonshotai/kimi-k3#providers)_

On InferenceX, we benchmark Kimi K3 serving performance directly on recorded internal claude code traces. We replay an hour of these traces as they reach a steady state. There is a median of 142k input tokens and a median of 444 output tokens per turn with a median of 65 turns per session. The short output tokens per turn is typical for workloads on agentic harnesses, where the agent calls tools frequently, even edits are tool uses.

This benchmark is a big step up from our previous 8k1k/1k1k benchmark, as it truly reflects real-world agentic use cases. From a systems perspective, it is also realistic and closest to production systems. It can reflect KV cache behavior, including prefix cache and KV offloading to DRAM.


_Source: [InferenceX](https://inferencex.semianalysis.com/datasets)_

For Kimi K3, Day 0 bringup was easier than DSv4 due to better documentation and preparation ahead of weights release. Appropriate images and a speculative decoder model were released at the same time as the weights.

For Nvidia, bringup was simple. But due to the models’ sheer size, it doesn’t fit on a single B200 node. We had to use PP to get it working. DSpark also didn’t work with PP.


_Source: [InferenceX](https://inferencex.semianalysis.com/)_

For B300, the model fits on 1 node and serves well. After accounting for the weights, GPU HBM can only hold 3.25M tok. In the graph below, throughput goes up as batch sizes increase until concurrency increases above 8. This roughly correlates to the 3.25M tok KV cache budget, and cache starts to thrash, resulting in hit rates falling to < 10% when theoretical hit rate is 95%.

But for a system with 219.91 GB/rank CPU DRAM offload, this results in an additional 15.76M tokens worth of KV cache, or a 4.85× total tokens worth of KV cache. And the performance gain is huge. With DRAM offload and without speculative decoding, This enables configs with concurrency above 8 to run without cache thrashing.

Cache thrashing configs appear in the bottom left of the throughput-interactivity graph, and are characterized by very low cache hit rates.


_Source: [InferenceX](http://Inferencex.com)_

## TCO

From SemiAnalysis claude code and codex usage, costs from cache read make up most of our spending.


_Source: SemiAnalysis_

At a P90 of 30 tok/s/user, serving with 1 node of B300 utilizing DSpark speculative decoding, gives you an input throughput of 4844.1 tok/s/GPU, translating to a cost of $0.1712/M input tokens for 3Y rental prices of $2.60/hr as of July 2026. This means providers are offering at an order of magnitude more than cost on serving Kimi K3. On B200 this is much worse and nearly infeasible.


_Source: InferenceX_

Using our AgentX dataset, which has 315:1 ISL to OSL ratio, we estimate the blended price of using Kimi K3 on Moonshot’s platform is $0.74 per million tokens at 22 tok/s/user. At a similar interactivity, our InferenceX results shows serving Kimi K3 on B300 with vLLM costs $0.171 per million tokens at more than double the interactivity, a 76.89% reduction in cost.


_Source: InferenceX_

### [E1799] Seeking Alpha | 2026-08-03 19:19 UTC | 盘中
AI infrastructure: Checking in on hyperscaler capex
URL: https://seekingalpha.com/news/4623520-ai-infrastructure-checking-in-on-hyperscaler-capex?utm_source=feed_news_all&utm_medium=referral&feed_item_type=news
Investors Should Know: Big Tech's AI infrastructure buildout has entered a new phase, with investors shifting focus from raw spending levels to whether capital expenditures are generating visible returns. Recent earnings from the four largest cloud operators show a widening gap between those hitting that bar and those still falling short.
Background
Four of the largest cloud and AI infrastructure operators, namely, Microsoft (MSFT), Alphabet (GOOGL), Meta Platforms (META), and Amazon (AMZN), have collectively committed hundreds of billions of dollars to AI data centers, chips, and compute capacity. Their Q2 earnings reports, released over the past couple weeks, brought that spending under intense scrutiny.
The central question driving analyst and investor sentiment is no longer whether to spend, but whether the spending is producing returns fast enough. The market has moved from rewarding aggressive investment to demanding evidence that revenue growth can outpace capital expenditure growth, a threshold only some operators have clearly crossed.
Stocks to Watch
The buildout extends well beyond the hyperscalers themselves. Advanced Micro Devices (AMD) and Nvidia (NVDA) are competing for rack-scale chip dominance inside these data centers. Semiconductor equipment makers, including ASML Holding (ASML), Lam Research (LRCX), and Tokyo Electron (TOELY) sit further up the supply chain, supplying the tools needed to manufacture the chips powering AI workloads. Memory chipmaker Micron (MU) has also been cited as a potential beneficiary as memory's share of total AI infrastructure spending grows.
Adjacent infrastructure players are also seeing a direct impact. SPX Technologies (SPXC) raised its data center capacity target, and Navitas Semiconductor (NVTS) projected AI infrastructure will represent more than a third of its sales by year-end. Power availability has emerged as a separate bottleneck, reshaping energy infrastructure investment alongside compute capacity.
Key Takeaways
Microsoft has become the benchmark for AI infrastructure balance, with Azure revenue surpassing $100 billion annually and the company guiding to positive free cash flow in fiscal 2027, even as capital expenditures rose roughly 70% year-over-year.
Alphabet and Meta face heavier investor skepticism, with both companies reporting negative or near-zero free cash flow while raising their forward capex outlooks.
Morgan Stanley analysts estimate hyperscaler AI infrastructure can generate 25–50% long-term returns on invested capital, with GPU rental businesses potentially producing 60–70% incremental operating margins. However, these returns remain forward-looking.
The buildout is reshaping adjacent sectors, including chip supply chains, semiconductor equipment, memory, and power infrastructure, as bottlenecks shift across the ecosystem.
News Roundup
Recent earnings and announcements from the major cloud operators and their chip partners illustrate how differently each company is navigating the tension between AI investment and financial discipline. The items below span quarterly results, analyst reactions, and strategic moves that together define the current state of AI infrastructure spending.
Microsoft (MSFT) — Azure revenue surpassed $100 billion annually for the first time, with quarterly growth accelerating four points sequentially to 43% year-over-year. BNP Paribas analyst Stefan Slowinski noted Microsoft had reached "an important step... in achieving an equilibrium state where revenue growth can ultimately outpace CapEx growth," while RBC Capital Markets, J.P. Morgan, and Morgan Stanley all reiterated bullish stances. The company generated $19.6 billion in free cash flow despite capital expenditures rising roughly 70% year-over-year to $41 billion, and guided to positive free cash flow in fiscal 2027.
Alphabet (GOOGL) — The company spent $44.9 billion on capital projects in a single quarter and raised its 2026 capex guidance to $195–205 billion, up from $180–190 billion previously, sending shares lower. Alphabet reported negative free cash flow of $5.9 billion for the quarter, even as cloud revenue grew 82% and contracted backlog reached $514 billion. Management acknowledged that leaning on third-party capacity as a bridging strategy creates "modest margin pressure," and the CFO confirmed the company remains supply-constrained even after the guidance increase.
Amazon (AMZN) — Following strong Q2 results, Amazon announced it would boost AI and technology spending by an additional $20 billion, with CEO Andy Jassy describing AWS as "booming" and noting its AI and chips businesses each surpassed $25 billion run rates. Ahead of the print, Wedbush had projected Amazon would raise 2026 capex guidance above $200 billion and 2027 guidance to $266 billion, with analysts watching Trainium chip capacity and free cash flow trends closely.
Meta Platforms (META) — Shares fell after earnings as investors reacted to AI capex expenses surging 55% quarter-over-quarter, free cash flow dropping to $784 million against $31.9 billion in operating cash flow, and the company taking on $24.9 billion in additional debt while suspending buybacks. Meta raised its 2026 capex outlook to a $137.5 billion midpoint, presenting what analysts described as the most polarized sentiment among the four hyperscalers.
AMD (AMD) — An expanded Azure partnership deploying AMD's Helios rack-scale platform, combining Instinct MI455X GPUs, EPYC Venice CPUs, and Pensando networking, gave a boost to AMD shares, positioning the company as Nvidia's most credible rack-scale competitor. At its Advancing AI 2026 event, AMD raised its long-term AI accelerator total addressable market forecast to roughly $1.4 trillion by 2030, alongside a CPU total addressable market raised to over $200 billion.
Nvidia (NVDA) — Nvidia has struck revenue-sharing deals with AI cloud partners to deploy multi-tenant AI factories, positioning the company within the broader $1.5 trillion in projected 2027 global cloud and AI capex that analysts are tracking.
SPX Technologies (SPXC) — The company lifted its data center capacity target to $1.1 billion from $750 million and raised full-year adjusted earnings per share guidance, reflecting direct demand from the AI infrastructure buildout.
Navitas Semiconductor (NVTS) — Navitas projected that AI infrastructure will represent more than a third of its sales by year-end, citing record book-to-bill as evidence of accelerating demand from data center customers.
AI Infrastructure Stocks in Focus
The screener below identifies publicly traded companies classified under Seeking Alpha's AI cloud and high-performance computing themes, as well as the AI chip supply chain, covering the compute capacity buildout and the semiconductor supply chain that supports it. Companies are ranked by relevance to these themes, with the most theme-aligned names appearing first.
The thematic screener draws primarily from the chip supply chain layer of the AI infrastructure ecosystem. Results include semiconductor equipment makers and chip manufacturers whose businesses are directly tied to producing the compute capacity at the center of the hyperscaler buildout. Market caps and Quant Ratings vary across the list, reflecting the range of company sizes operating across this supply chain.
Companies in the screener include Tokyo Electron Limited (TOELY), ASM International NV (ASMIY), SUSS MicroTec SE (SESMF), MKS Instruments (MKSI), Lam Research Corporation (LRCX), QuickLogic Corporation (QUIK), Photronics (PLAB), Atomera Incorporated (ATOM), ASML Holding N.V. (ASML), and Onto Innovation (ONTO).
Explore Further
Stock Screeners — Browse and customize predefined screens for AI infrastructure and technology themes.

### [E1873] 华尔街见闻快讯 | 2026-08-03 19:50 UTC | 盘中
亚马逊正式跻身3万亿美元俱乐部！AWS负责人：“与客户签五年期长期合同，需求远超供给”
URL: https://wallstreetcn.com/articles/3778603
亚马逊继上周五大涨15%后，周一盘中再次收涨4.6%，市值突破3万亿，成为史上第五家跻身这一量级的企业。亚马逊AWS负责人Matt Garman周一表示：“目前，市场需求仍远超供给，正加快建设和投资步伐，正在与客户签署五年期承诺合同。”

## 研究交付要求

这是一条已经从整场材料语义拆出的独立催化，不要重新把它压回大盘总评。核实美股异动原因和传播时序；用结构化供应链、内部 KB 与公司披露坐实每个亚洲候选；逐腿查真实行情和本地市场 price-in；候选不成立也要逐条给出有证据的 AVOID。