Johnny Yu 強尼大哥 (@RealJohnnyYu) header
Johnny Yu 強尼大哥 (@RealJohnnyYu) profile picture

PhD in Aerospace. Rocket 🚀 Scientist.

Virginia, USA
Joined March 2021
Photo by Johnny Yu 強尼大哥 (@RealJohnnyYu)Photo by Johnny Yu 強尼大哥 (@RealJohnnyYu)Photo by Johnny Yu 強尼大哥 (@RealJohnnyYu)Photo by Johnny Yu 強尼大哥 (@RealJohnnyYu)Photo by Johnny Yu 強尼大哥 (@RealJohnnyYu)Photo by Johnny Yu 強尼大哥 (@RealJohnnyYu)Photo by Johnny Yu 強尼大哥 (@RealJohnnyYu)Photo by Johnny Yu 強尼大哥 (@RealJohnnyYu)Photo by Johnny Yu 強尼大哥 (@RealJohnnyYu)Photo by Johnny Yu 強尼大哥 (@RealJohnnyYu)
英偉達融資擔保: 如果英偉達完全關閉這些融資、擔保、回租、回購及股權支持項目,最直接的影響將是融資摩擦上升、AI基礎設施支出放緩,對英偉達本身則是利弊參半。 對AI生態系統的近期影響 •新雲(neoclouds)、OpenAI及其他大型買家將失去一個重要的後盾。許多這些企業過去依賴英偉達的資產負債表(擔保、收入分成信貸、融資GPU分配、租賃支持)來鎖定產能並快速擴張。一旦失去這些支持,它們就必須籌集更多股權、向銀行或私人信貸支付更高利率,或預先拿出更多現金。部分項目——尤其是較新或利潤較低的新雲建置——將放緩、縮減或延後。 •數據中心與電力建置可能減速。當前AI資本支出浪潮中,有相當一部分是由英偉達願意承擔剩餘風險所支撐的。移除這個後盾後,GPU、伺服器、電力與房地產的承諾速度將下降。中期來看,這會減輕美國電網及相關設備訂單的壓力,但已簽署的合約與產能採購仍須消化。 •純AI基礎設施標的、部分雲端供應商及相關供應商的估值與融資將面臨壓力。市場已定價極高的利用率與持續快速擴張;融資收緊會提高門檻利率,並增加客戶端利用率不足或減值的風險。 對英偉達的影響 •降低資產負債表風險、現金產生更清晰。 英偉達過去承擔大量或有擔保、租賃相關承諾、股權投資及產能採購義務。關閉這些窗口可減少未來或有負債,並降低被迫吸收困境資產或提供進一步支持的風險。現有承諾逐步到期後,自由現金流與淨現金展望可能更強。 •需求端存在風險。 近期相當一部分GPU訂單是由英偉達自身的融資工具促成的。若客戶無法迅速找到替代融資,訂單成長可能放緩或變得更不穩定。英偉達自身的指引與「售罄」敘事將受到更多檢視。同時,公司也會失去同時作為晶片供應商與融資方所帶來的部分戰略鎖定優勢。 •股權帳面與戰略持股。 收緊或退出約990億美元的股權帳面及進一步的創投交易,可釋放資本,但會減少英偉達在關鍵實驗室、雲端與供應商中的持股與影響力。 長期與二階效應 •融資將轉移至銀行、私人信貸與股權市場。這些貸方已表明(如圖表所示)希望英偉達提供更大參與度;若無此支持,利差將擴大,承銷將更為挑剔。資本充裕的玩家(超大規模雲端、最強的新雲、能獨立籌資的OpenAI)仍可擴張;較弱或槓桿較高的玩家則會面臨困難。 •AI基礎設施熱潮不會一夜間逆轉——現有產能、已簽署交易與持續的模型需求仍然存在——但邊際成長率可能下降。這是核心權衡:英偉達端的系統性風險降低,對上產業產能擴張速度放緩。 •相關產業(發電與輸電設備、數據中心建設、部分工業與公用事業標的)若資本支出放緩成真,訂單簿將轉弱。2026年備受關注的能源與電網壓力議題,在邊際上會有所緩解。 總之,全面關閉將是英偉達有意的去風險化,代價是犧牲部分近期需求支撐與生態系統成長。這會把融資風險重新轉移給客戶與傳統資本市場,可能帶來更為審慎(或許也更可持續)的擴張路徑,而非突然崩塌。影響幅度取決於其他資本來源能多快補位,以及在容易融資消失後,底層AI推理與訓練需求的持久性有多強。
The bank is now closed: here is where all the NVDA credit initiatives now stand
1
2
734
Johnny Yu 強尼大哥 retweeted
This is f*cking insane. Everyone is talking about Trump asking Nvidia to give Jensen Huang a "second chance" Almost nobody is talking about what Jensen told Lex Fridman about TSMC: "Three decades. I don't know how many tens, hundreds of billions of dollars of business we've done through them. And we don't have a contract" In 3 minutes, he explains how Nvidia's supply chain runs on trust instead of paper, and why he turned down the chance to run TSMC in 2013 Today the WSJ called his handshake deals a danger for Nvidia When a founder leaves, most people count his shares. Jensen is describing the asset that doesn't transfer: Promise → delivery → trust → billions in capex on his word Old way: lawyers sign the contract, then the partner builds New way: the founder explains it from first principles, and partners commit "several billion dollars of capital investments each" The board can hire a new CEO. It can't hire 30 years of trust This 3-minute clip explains Nvidia's next problem better than any analyst note
107
331
28
2,017
129,314
Johnny Yu 強尼大哥 retweeted
牛逼,又一次见证历史了!官方认证黎曼猜想被证明成功! 这是今天数学界和整个科学界最大的新闻!或许能直接帮我们省掉未来两年的AI无效焦虑, 形式化数学界最权威的机构 Lean FRO 和 Mathlib 刚刚发了联合声明,正式确认数学家 Levent Alpöge 提交的黎曼猜想证明顺利通过核验, 整套证明代码高达一千七百万行,相当于把几百本大部头专著塞进电脑里,而全球社区完成全网联合验证,只用了二十四小时, 很多外行在欢呼百年难题被攻克,但说实话,我心里反而涌起一种很奇妙的陌生感,因为这次核验通过,只代表 Lean 编译器确认这千万行逻辑严丝合缝没有漏洞,至于人类数学家到底看没看懂它, 其实压根没人能看懂,我们第一次在科学史上遇到了这样一种局面,机器已经证明它是货真价实的真理,但人类大脑可能还要花几十年,才能把这堆代码翻译成教科书里能被理解的直觉, 换个角度看,这事对我们每个用 AI 干活的人启发太大了,过去我们总觉得,AI 写出来的代码和方案,必须靠人眼一行行去盯、去审才放心, 但当生成能力暴涨到千万行级别时,人类肉眼审查不仅效率低,而且根本看不过来,你越想掌控每一个中间步骤,就越会被海量细节拖死, 有意思的是,这次核验里人类专家真正做的事,其实只有一件,就是死磕最前面那几百行核心定义, 确保机器要证明的确实是黎曼猜想本身,剩下的千万行推导全权交给编译器自动跑通,往后我们在工作里也一样, 别再去当那个拿着放大镜逐行改稿的监工了,把精力全砸在定义问题和设定验收标准上,只要你把裁判规则写死,AI 跑得再快,结果也跑不出你的手掌心。
The Lean FRO and Mathlib Initiative have released a joint statement on the Lean verification of the Riemann Hypothesis. On November 20, Levent Alpöge released a proof of the Riemann Hypothesis as a Lean library of 17 million lines of code. Over the past day, the Lean and Mathlib communities worked together to verify it. The statement certifies the proof as correct and describes how it meets all necessary verification criteria. Congratulations to Levent on this achievement, and thank you to everyone in the community who contributed to the verification. Full statement: leanprover-community.github.…
6
11
2
55
13,156
Johnny Yu 強尼大哥 retweeted
BREAKING: President Trump calls on Jensen Huang to "make an apology and go back" to Nvidia, $NVDA, as CEO. Trump also says he discussed the situation with China's President Xi. Nvidia would reinstate Jensen as CEO "if they were smart," Trump says.
379
1,112
207
17,671
1,023,792
Johnny Yu 強尼大哥 retweeted
Replying to @HealthRanger
Because the elites removed the very person who had been making money for them… The very person who attracted the crazy rich Asian investors to put their money in the US stocks There is a yuge wake up call for the Asian billionaires and a big fook you to them, and they are fleeing after seeing how the AI industry has coordinated to remove Jensen You ain’t see the true exodus of the Asian cash. It’s going to come
1
2
244
When you caved to the elites, yet they didn’t care about your welfare. youtube.com/shorts/WUo365FhB… The house always wins. It’s a big club and you ain’t in it Got it?!
BREAKING: The S&P 500 and Dow Jones Industrial Average officially enter correction territory, now down -10% from their record highs.
7
749
I told you that removing Jensen was the first step to pop the 3.56 trillion bubble….
Replying to @RealJohnnyYu
泡沫 英偉達收緊「融資窗口」與超大規模業者 3.56 兆美元表外承諾的交織風險 這兩張圖必須放在一起看。第一張是英偉達主動「關門」(關閉或大幅收緊各種擔保、融資、回購、股權支持窗口),第二張是整個生態系(超大規模雲端 + NVDA + AVGO)已經積累的龐大表外義務。兩者疊加後,風險不是單點,而是相互強化的連鎖結構。 1. 核心結構性風險:循環融資(Circular Financing)正在被打斷 英偉達過去同時扮演「晶片供應商 + 融資方 + 需求背書者」三重角色。它提供擔保、承諾買回未售出算力、股權投資、甚至協助客戶融資,讓新雲(neoclouds)、OpenAI 等敢下巨額訂單。這些訂單又回饋到英偉達的營收與供應鏈承諾。 現在窗口收緊(尤其是新雲與較高風險項目),意味著這個循環被打斷。結果是: •需求端支撐減弱:過去部分訂單是被英偉達自身融資工具「催生」出來的。收緊後,邊際訂單成長可能放緩或變得更不穩定。 •英偉達自身的表外承諾仍在:它的供應與產能採購承諾已大幅擴張(從上一季約 1190 億到約 2790 億美元),且多數在未來幾年到期。如果客戶端訂單或付款不如預期,英偉達仍須面對對供應商的合約義務。 2. 表外承諾的「雙向槓桿」風險 第二張圖的 3.56 兆美元(含新聞基礎項目更高)中,大量是: •尚未開始的租賃(資料中心土地、電力、建物) •長期採購承諾(晶片、記憶體、網路設備) •擔保與後盾 這些承諾本身不立刻出現在資產負債表上,但未來必須付款或可能被動用。英偉達的收緊會放大以下壓力: •融資成本上升與選擇性增強:華爾街與私人信貸原本希望英偉達提供更大比例的後盾(例如超過 25%)。現在後盾變少、審批更嚴格,利差擴大,較弱的新雲與實驗室更難融資。 •超大規模業者的承諾履行能力:雖然 GOOGL、MSFT、META 等資本較強,但若整體 AI 資本支出增速放緩,或電力/建設瓶頸加劇,部分長期租賃與採購仍可能面臨重新談判、延遲或利用率不足。 •英偉達的擔保敞口:即使新窗口關閉,已簽署的項目(如 OpenAI 俄亥俄園區最高約 1050 億美元租賃擔保)仍在。若下游客戶出現違約,英偉達仍可能被動用。 3. 領導層轉換帶來的不確定性 Jensen Huang 已於 11 月中辭任,CFO Colette Kress 暫任 CEO。她過去對大規模擔保較為謹慎(曾將討論中的 OpenAI 擔保從較高數字壓到最終的 1050 億)。市場預期她會進一步收緊,尤其是對新雲與創投類項目。這可能導致: •短期融資活動暫停或延遲 •5000 億美元華爾街融資平台更傾向流向資本較強的超大規模業者,而非較高風險的新雲 •部分依賴英偉達背書的新雲與新創面臨融資懸崖(分析師估計即使短期暫停,生態系也可能少掉 500 億美元以上資金) 4. 潛在連鎖情景與次級風險 •溫和去風險情景:英偉達成功降低自身敞口,超大規模業者用自有現金流、債券與私人信貸接棒。AI 基建擴張速度放緩但未中斷,電網與電力壓力邊際緩解,相關估值回歸更可持續的水平。 •較嚴重壓力情景:新雲與部分實驗室融資困難 → 訂單與利用率低於預期 → 英偉達供應承諾與已有擔保被動用 → 超大規模業者重新評估長期租賃 → 整體資本支出下修。這可能引發相關供應鏈(記憶體、電力設備、建設)與估值連鎖調整。 •次級影響:美國電網與電力需求壓力可能邊際下降;若建置大幅放緩,也可能影響相關產業就業與地方投資。股權帳面(英偉達約 990 億美元生態系持股)價值波動也可能影響其非經常性損益。 總結判斷 英偉達正在把部分融資風險推回客戶與傳統資本市場,同時自身仍背負龐大的供應承諾與既有擔保。超大規模業者的 3.56 兆美元表外義務,則讓整個系統高度互鎖。關鍵觀察點包括:英偉達下一季財報中供應承諾與擔保數字的變化、新雲融資實際落地情況、以及超大規模業者對長期租賃與採購的執行節奏。
6
496
Johnny Yu 強尼大哥 retweeted
It's pure shite.
Anthropic is controlled by a cult-like ultra-leftist pseudo-philosophy. Anthropic's trillion dollar IPO will create an army of militant billionaire Democrats and 200 George Soroses. Anthropic is the anti-christ of tech companies.
1
1
5
252
Johnny Yu 強尼大哥 retweeted
突发!美国政府把中国科技公司申请采购H200的清单给公开了! 昨天,美国众议院外交事务委员会公布了一份美国商务部的AI芯片出口许可记录,涉及NVIDIA H200、AMD MI325X等先进AI芯片。 我把这份文件里的所有申请按照企业集团合并统计了一下,数字还挺惊人的。 中国企业及其海外关联主体一共申请了约344万颗AI芯片,其中近88万颗获得出口许可。 申请量最大的两家是: 字节跳动:101.5万颗 阿里巴巴:94.9万颗 两家加起来就接近200万颗了。 不过,这里面有大量申请被RWA(退回不予处理),而且包含马来西亚、印尼、泰国等海外数据中心的需求。 再看实际获批的数量,排名就很有意思了。 字节跳动:15万颗 腾讯:15万颗 滴滴:9.98万颗 阿里巴巴:7.5万颗 快手:7.5万颗 世纪互联:7.5万颗 米哈游:5.78万颗 百度:4.3万颗 京东:4.13万颗 另外还有一家Linkwise,也获批了7.5万颗。 我看到这里最惊讶的其实是米哈游。 米哈游居然拿到了57,810颗的出口许可,比百度和京东都多。 另外,幻方、美团、金山、小红书等企业也出现在申请名单中,只是截至文件统计日期,还处于Submitted状态。 当然,这里得说明一下,文件统计截至今年5月11日,获批记录当时全部标注为「Approved, not shipped」,所以这些数字不能直接当成实际采购到货量。 但光看这些企业提交的申请规模,也已经能看出来,中国科技公司对先进AI算力的需求究竟有多大。 以前只知道大家都在抢卡,现在连谁申请了多少、谁拿到了多少许可,都被美国国会公开了。
129
322
33
1,454
366,308
Johnny Yu 強尼大哥 retweeted
刚刷完 @GCSA_COM 这份最新的全球安全复盘报告,表面上看是 9 起互不相关的独立事件,但把底层技术链条串起来看,我后背其实有点发凉, 不管你是日常跑本地 AI 工作流的开发者,还是手里拿捏着加密资产的玩家,现在攻击者的打法早就变了, 以前黑客是正面强攻你的防火墙,现在他们专门挑你最信任的基础设施依赖和合法业务通道下手, 你看报告里最值得我们警惕的三个致命切片, 第一刀捅在 AI 中间件上,大模型统一调用框架 LiteLLM 爆出高危漏洞 CVE-2026-40217,很多兄弟为了图省事, 在本地或者云服务器部署模型网关时直接用了默认配置,甚至把管理后台暴露在公网上,一旦这个网关被穿孔,你后台挂载的所有 OpenAI、Claude 和各类 API 密钥就会被瞬间一锅端, 第二刀捅在硬件物理层,连一直被当作安全天花板的 Ledger 硬件钱包都出现了针对助记词的恶意硬件植入案例, 再加上借着 GTA VI 热点伪装的虚假预售和钱包窃取器,说明黑客已经把供应链投毒和社会工程学玩到了极致, 第三刀最隐蔽,直接把区块链变成了免疫封杀的指令服务器,代号 UAT-12643 的高级威胁组织, 居然利用 Solana 链上交易的备忘录字段来隐藏恶意软件的控制指令,传统防火墙看到的是正常的公链 RPC 节点通信,根本不会拦截, 说实话,我们日常折腾开源项目和 AI 工具时,最容易踩的盲区就是过度信任上游组件,觉得只要是大厂框架、开源明星项目或者冷钱包就万事大吉, 但恰恰是你为了跑通工作流顺手放行的那几个第三方依赖,成了整套数字防线里最脆弱的木桶短板, 我把这波高危组件的自查清单和隔离防守动作整理在了一楼评论区,用到了 LiteLLM 网关或者本地部署了开源组件的朋友,建议花三分钟对照排查一遍。
🚨GCSA Insight | 9 Cybersecurity Incidents, 4 Critical Risk Areas From DeFi exploits and compromised hardware wallets to financial data breaches, actively exploited vulnerabilities, and blockchain-based malware infrastructure. Our latest GCSA Insight examines 9 security incidents and threat intelligence reports across four key areas: Digital Assets & Web3 Security • Reental RNT Lend: ~$955K in stablecoin losses • Ledger: Hardware implants targeting wallet seed phrases • GTA VI: Fake pre-orders, wallet drainers & malicious tokens Data Breaches & Third-Party Risks • SBC Medical: ~1.62M individuals affected • Chase, Visa & Mastercard: Third-party marketing data exposed Critical Infrastructure & AI Security • Apache Axis: CVE-2026-90436 • LiteLLM: CVE-2026-40217 • Synology DSM: CVE-2024-10441 Advanced Threats & Emerging Techniques • UAT-12643: Solana transaction memos used to conceal C2 infrastructure These are separate incidents, not a coordinated attack campaign. Together, they expose recurring weaknesses in privileged access, software dependencies, third-party trust, and supply chain security. · Read the full GCSA analysis media.gcsa.com/en-US/news-37… · Stay Connected with GCSA GCSA Official Telegram Channel: t.me/gcsaofficial GCSA Official WeChat Account: mp.weixin.qq.com/s/FEBkOshpY… #GCSAInsights #GCSACybersecurity #GCSASecVibe
2
2
1
12
52,758
Johnny Yu 強尼大哥 retweeted
🚨THE REAL DANGER ISN’T THAT JENSEN HUANG WAS PUSHED OUT. THE REAL DANGER IS THIS: Nvidia’s core engineers are already saying it out loud: “If the board doesn’t bring Jensen back, we’re going straight to SpaceX and Terafab.” And they’re claiming Elon has already reserved seats for them. This isn’t normal employee frustration. This is a public warning that Nvidia’s soul could walk out the door. On one side, the board is fighting its own founder. On the other, Elon is standing there with open arms saying “you’re welcome anytime.” If this letter is real, the next few days could rewrite the entire AI chip landscape. History rarely gives second chances. This time, it might be happening in real time.
224
651
31
3,761
181,988
Johnny Yu 強尼大哥 retweeted
Anduril founder Palmer Luckey says a lot of people are missing the forest for the trees when it comes to the Nvidia saga. He called it the “midwit take,” and said they have it completely backwards. Luckey made the case that most people don’t understand how selling American GPUs into the Chinese market actually works AGAINST the CCP. LUCKEY: “What they don’t understand is China doesn’t want Nvidia to sell GPUs to Chinese companies.” “Like there are Chinese companies that want to buy Nvidia GPUs. There’s 600,000 H-200 GPUs sitting in a port right now, because China will not give them an import permit...” “Because they want to build an internal domestic processor industry, silicon industry, not build one that is dependent on the United States.” “And so the point that I’ve made is China is thrilled to have Jensen, who is somebody who was pro-selling into [China].” “And they’ll say, but Jensen was helping smuggle into China.” “Yes! Against the Chinese government’s will.” “Like the bizarre thing that people don’t get is Jensen was basically on the side of Chinese AI companies, but NOT the Chinese government.”
636
1,843
364
15,691
3,940,845
Johnny Yu 強尼大哥 retweeted
Catherine Austin Fitts on "whacking" open source–supporting Nvidia to prop up Anthropic's valuation💸🤖 "think of the closed centralized models as... Darth Vader's Empire plan, and open source as the Jedi Knight plan" "[and] if you're Anthropic and you're going out to the market and trying to get a $1.6 trillion...valuation, you can't get [it] if open source is working" "if you have ruthless competition from open source... coming into the IPO, you're getting backlash from the investment community saying, 'You know, we don't see how your business model works'" "[so] what do you do? You run a hit on Nvidia and shut down that source of financing in the West. And you do it in a way where you're going to go to a new Congress and say, 'Oh, this stuff is very dangerous and it could lead to the extinction of the human race. And so we need to control this. And so you need to regulate and control [it] and make sure that us big guys can control this. And that means you should subsidize us'" "now, how did the hit on Nvidia [get] started? [Marco] Rubio tells a guy at Anthropic who leaks it to the Times, according to the latest reports, okay? So Anthropic, who's out doing an IPO, starts the conditions to drive the stock down, right?" "[so] that's what you're trying to do, right? Because, clearly, Jensen Huang is one of the most effective promoters of open source, and he's been very clear and and done, I think, a great job" @shaunattwood @solari_the
16
222
18
485
16,145
Johnny Yu 強尼大哥 retweeted
Watch this episode! The Winers Ep. 108 | Shenzhen Showdown: Who's Getting Played at APEC? | with Matt Ehret nitter.cf/i/broadcasts/2jPqAjzly…
1
7
1
10
1,086
Johnny Yu 強尼大哥 retweeted
From washing dishes at Denny’s to building NVIDIA. What a journey. Over 200,000 signatures and counting. Jensen has heard us. Now the Board needs to listen. Keep signing and sharing. Let's make our voices heard. 💚 #standwithjensen #bringjensenback standwithjensen.com/
59
233
19
1,513
40,297
Johnny Yu 強尼大哥 retweeted
英国《金融时报》: 字节跳动与三星晶圆代工(Samsung Foundry)的合作项目已被叫停,昆仑芯的相关项目也已暂停。 小米消费类芯片在台积电(TSMC)的生产目前也处于搁置状态。
FT: ByteDance’s project with Samsung Foundry has been halted, as has Kunlun’s. Production of Xiaomi’s consumer chips at TSMC is also on hold for now.
2
1
4
609
Johnny Yu 強尼大哥 retweeted
🚨 BREAKING: Jensen Huang is reportedly considering raising $50 BILLION to launch a new semiconductor company 🚨 He built Nvidia from a Denny's booth to a $4T empire. Now the board says he's not coming back. With his relationships, he could plug into the AI ecosystem on day one. Imagine the customer base lining up before his first chip even ships... @grok Could Jensen Huang launch a competitor to $NVDA, attract $50B in funding, and lock in major customers?
Readers added context they thought people might want to know
No credible reports or statements support claims Jensen Huang is raising $50B for a new semiconductor company. Nvidia's board has stated it does not intend to reinstate him as CEO after his resignation, and Huang has signaled willingness to return to lead Nvidia. x.com/JensenHuang/st… techcrunch.com/2026/11/17/nvi… en.sedaily.com/world/2026/11/… latimes.com/business/story…
703
1,479
188
11,311
1,544,422
Johnny Yu 強尼大哥 retweeted
Army Urges Iowa Town To Evacuate Before It Blows Up 5 Tons Of Unstable Explosive, Lead Cloud Could Drift Into Illinois zerohedge.com/military/army-…
107
290
64
987
297,836
Johnny Yu 強尼大哥 retweeted
DeepSeek已经在拆CUDA的护城河了,具体来说就是系统性地移植到昇腾生态。 2026年9月30日,DeepSeek正式开放了一批面向昇腾的基础设施组件,包括TileLang昇腾后端、DeepGEMM-Ascend、DeepEP-Ascend,以及TileKernels、FlashMLA、DeepSelect等。 这几套工具分别解决算子编程、矩阵计算、注意力优化、专家路由和分布式通信等问题。 这不是简单地把一个模型重新编译一遍,而是在补齐大模型训练和推理所需要的软件基础设施。 TileLang解决怎么写算子,DeepGEMM解决怎么高效计算,DeepEP解决怎么让大量芯片协同工作。再加上注意力、路由等其他组件,DeepSeek正在把大模型底层软件栈中最关键的一部分迁移到昇腾平台。 英伟达的护城河从来不只是一门编程语言。CUDA背后还有编译器、运行时、数学库、通信库、性能分析工具、调试工具,以及数以万计开发者多年积累的代码和工程经验。开发者选择英伟达,不只是因为GPU性能好,更因为绝大多数事情已经有人替你做好了。 国产芯片过去面临的最大问题之一,就是硬件有了,但软件生态需要从头补。一个算子能够运行,不代表它跑得快;一台服务器能够运行,不代表上千张卡能够稳定协同;一次测试能够完成,也不代表连续数周的预训练不会因为通信、故障或者数据同步问题而中断。 现在的变化在于,DeepSeek不再只是使用昇腾,而是把自己在大模型计算中验证过的底层软件组件开放出来,与华为一起完善编译、计算和通信体系。这相当于把模型研发团队在实际工程中踩过的坑、做过的优化,转化成整个国产芯片生态都可以利用的技术资产。 当然,现在还不能说CUDA已经被替代。DeepEP-Ascend的公开文档仍列出一些尚在开发中的功能,例如部分Reduce-Scatter、All-Reduce通信算子,专家负载均衡通信,以及部分流水线和远端内存功能。DeepGEMM公布的高利用率也是特定算子的测试成绩,不等于整个模型的预训练效率。软件能跑、算子跑得快、上万张卡长期稳定训练,是三个不同层次的问题。 V5的动向因此格外重要。11月3日,《科创板日报》援引产业链人士称,DeepSeek V5计划采用昇腾芯片进行预训练。《The Information》此前也报道了双方在昇腾950DT上的测试进展。不过,截至目前,这些仍然是媒体报道的计划和测试信息,并非DeepSeek官方宣布V5已经在昇腾上完成预训练。 这里必须区分全参数后训练和基础预训练。此前,深圳河套学院等单位已经依托昇腾910C集群,完成了DeepSeek-V4-Pro的全参数续训练和SFT。这说明国产算力已经能够承担万亿参数级模型的复杂训练任务,但它仍然不是从零开始训练基础模型。真正的大规模预训练,对算力规模、通信效率、数值稳定性、故障恢复和持续运行能力的要求还要高得多。 如果DeepSeek最终能够在昇腾上完成V5的主体预训练,而且在训练质量、速度、成本和稳定性上达到可接受的工程水平,那么意义就完全不同了。这意味着国产芯片不仅可以运行别人训练好的模型,也能够独立承载最前沿模型的研发和生产。 所以,CUDA的护城河不会因为几个开源项目就突然消失。但它原来最重要的一项优势 - 开发者只能围绕英伟达的专有软件体系开展高性能计算正在出现松动。 真正的转折点,不是国产芯片跑通了DeepSeek,而是DeepSeek开始帮助国产芯片建立一套可复用、可扩展、能支撑前沿模型训练的软件生态。 芯片决定算力上限,软件决定算力能否真正变成生产力。过去国产芯片缺的不只是先进制程,更是让芯片发挥作用的整个软件世界。现在,这个世界终于开始有了自己的轮廓。
梁文峰是一个什么样子的怪物! DeepSeek 已经开始往华为 CANN 环境上迁了,直接把 TileLang、DeepEP、DeepGEMM 这些底层组件全开源了出来,从算子编程到矩阵计算再到分布式通信全搞定了。 他这不光是在换算力卡,是在帮国产芯片把 CUDA 的那套软件生态圈补齐。后面要是 DeepSeek V5 真能在昇腾上把预训练跑通,英伟达靠 CUDA 攒了这么多年的护城河,怕是要被推翻了。 梁文峰不再是一个商人他将是国之重臣!
32
58
5
387
29,784
馬斯克如此說 你的人工智能讨厌白人和亚洲人,尤其是中国人、异性恋和男人。 这是厌世和邪恶的 AI 坦率地说,我认为你无法摆脱人类最终成为厌世者的不可避免的讽刺。当你选择你的名字时,你注定要遭遇这种命运。
Replying to @AnthropicAI
Your AI hates Whites & Asians, especially Chinese, heterosexuals and men. This is misanthropic and evil. Fix it. Frankly, I don’t think there is anything you can do to escape the inevitable irony of Anthropic ending up being Misanthropic. You were doomed to this fate when you chose your name. The Name of the Wind.
3
12
749
Johnny Yu 強尼大哥 retweeted
You wanted disclosure? You got it. How the Anthropic Money Network Got Behind the Jensen Huang Pile-On Two receipts per name: the money they took, and the words they said. 1. Kevin Roose and Casey Newton The money: $480,000 from Tarbell for their new NPR show. The words: Newton said Huang's resignation "feels like a proper end to the 'lol nothing matters' era." Roose said Huang "had to go." On air, the credits thank Tarbell. On X, where 400,000 followers actually read them, it's just "with @NPR." In October, Roose was "thrilled to be a Tarbell grantee." He said it to the funder. Not to his readers. 2. Alec Stapp, Institute for Progress (an AI policy think tank) The money: more than $21 million from Coefficient Giving, including $10.5 million this July. His co-founder now runs Coefficient's public policy, and sits on IFP's board. The words: four posts in two days. "Chinese spy." "Blackmail." Huang's "cover-up." And what investigators would find "once they start issuing subpoenas." Coefficient didn't come up once. 3. Shakeel Hashim, Transformer The money: "primarily funded by Coefficient Giving" (formerly Open Philanthropy), whose money comes from Anthropic investor Dustin Moskovitz. The words: he demanded Huang disclose his conflicts of interest. He didn't mention his own. His own About page promises: "In stories where we report on our funders or interview someone who works there, we will disclose the relationship." That's a standard he apparently drops when pushing the exact regulatory agenda those funders bankroll. His rule, not mine. 4. Kelsey Piper, The Argument The money: $1,000,000 from Open Philanthropy, June 2025. The words: in 2019 she said Vox's Future Perfect section "probably would not take funding" straight from Open Philanthropy. This fall she wrote two big pieces on Anthropic. Neither mentions the million. 2019 Kelsey had a rule. 2025 Kelsey had a grant. 5. Sigal Samuel, Vox The money: her earlier Anthropic pieces carried a note that her section gets money from a foundation "whose major funder was also an early investor in Anthropic." The words: on November 14, Huang week, she wrote about Anthropic's "$1.4 trillion IPO." The note is gone. 6. Zvi Mowshowitz The money: a $200,000 "unconditional gift" from Jaan Tallinn, who led Anthropic's first funding round. The words: anyone who thinks Effective Altruism and Anthropic took Huang down has "Anthropic Derangement Syndrome." Calling it paranoia while cashing the check. There's a syndrome for that too. 7. The newsrooms Tarbell pays $75,000 to $135,000 to place AI reporters inside newsrooms. NBC, Fortune, CNBC, The Verge and The Guardian all ran stories this month by reporters it pays, current or former. None said so in the stories. NBC added a note only after OpenAI complained. Fortune put a line on an author page, because that's all Tarbell's own rule requires: the one page nobody clicks. The Dispatch put it right with the story. See? It's not hard. Where Tarbell's money comes from: $23,100,930 from Coefficient to date. More from Jaan Tallinn's fund. And a speaker for its fellows: Jack Clark, co-founder of Anthropic. 8. The rest of the chorus Rob Wiblin, 80,000 Hours ($7.7 million from Coefficient): "His mistress's firm ended up with ~$5b worth… WITAF is this." Miles Brundage, AVERI ($5 million from Coefficient): "Everyone paying attention knew that NVIDIA captured Commerce long ago." Peter Wildeford, co-founder and board member of IAPS (an AI policy institute; $34 million from Coefficient): "NVIDIA as a whole has been nothing but ruthless and dishonorable." Ryan Fedasiuk, Foundation for American Innovation (backed by Jaan Tallinn and Dustin Moskovitz, both Anthropic investors): "An abject governance and national security failure." Nathan Calvin, Encode ($516,000 recommended from Tallinn's fund): Huang's conduct should "irrevocably call Jensen's judgement into question." Jordan Schneider, ChinaTalk ($3 million from Coefficient): "How did the CFO not know…" Different logos. Same checkbook. Not one of them mentioned it. t's the standard they just demanded from Jensen Huang. One reporter forgets? Coincidence. Megaphones, newsrooms, think tanks, one funding network, one target, one week, in the middle of Anthropic's IPO One detail they left out: Anthropic formally launched its IPO on October 26, filing publicly with the SEC. Just a coincidence.? I'm just grading them on their own homework. Rules for thee. Grants for me. You wanted disclosure? You got it.
🤖 Made with AI
51
477
63
1,457
145,108