The Newsroom: NVIDIA’s $35 Billion China Chip Orders, Salesforce's Braze Deal & Microsoft's Agent Controls
The Newsroom · October 21, 2026
China's Reported $35 Billion Nvidia Order, OVHcloud's Spare Capacity and a Weak Elice Debut Show Where AI Demand, Supply and Valuations Collide — Plus Microsoft's Agent Controls, Salesforce's Braze Deal and a BeyondTrust Zero-Day
Demand for AI Compute Is Not the Problem. Supply, Power and Valuation Are.
Three developments frame today's briefing. ByteDance, Alibaba and Tencent have reportedly ordered more than three million Nvidia RTX PRO 5500 cards, a sticker value above $35 billion, for an export-compliant chip that Nvidia reportedly cannot build in anything like that volume. OVHcloud says 110 MW of its connected capacity is open for new customers and plans its first AI model, a reminder that spare capacity is a commercial problem as well as a technical one. And Elice, a Korean AI cloud provider, fell about 32% on its debut despite an institutional book roughly 389 times covered.
Elsewhere, Microsoft announced controls that treat AI agents as insider risks, Goldman Sachs warned that power delays could keep roughly half of planned data center capacity from arriving on schedule, Salesforce agreed to buy Braze for about $5.66 billion, and a critical BeyondTrust zero-day is the week's most urgent patching item.
Dating note: Most stories were published on October 20, and some carry over from October 19 or earlier. Each item states its original date and whether it is a confirmed announcement or a media report. Yesterday's AI infrastructure coverage is in AI Watch (October 19); this briefing adds only what is new.
The Newsroom at a Glance
| Story | Date | Status |
|---|---|---|
| Chinese tech giants' Nvidia RTX PRO 5500 orders (>$35B sticker) | Oct 20 | Reuters exclusive; anonymous sources |
| OVHcloud: 110 MW free, first AI model planned | Oct 20 | CEO statements; model date unsettled |
| Elice falls ~32% on Kosdaq debut | Oct 20 | Confirmed market event |
| Microsoft: Guardian Agents and dynamic permissions for AI agents | Oct 19 | Company announcement; no availability or pricing given |
| Modal adopts Nvidia Vera CPUs for agent hosting | Oct 19 | Customer announcement; benchmarks are Modal's own |
| Goldman: power delays could slow data center openings | Oct 20 | Analyst estimate |
| BeyondTrust critical zero-day exploited | Oct 15 to 16 | Patch available; earlier story still active |
| Salesforce to acquire Braze (~$5.66B) | Oct 20 | Definitive agreement; pending approvals |
Top Technology Headlines
Chinese Tech Giants Reportedly Place More Than $35 Billion in Nvidia Chip Orders
What happened (October 20, media report): Reuters reported exclusively, citing two people familiar with the matter, that ByteDance, Alibaba and Tencent have ordered more than three million of Nvidia's new RTX PRO 5500 cards, with a combined sticker price above $35 billion, at roughly $11,400 per card. The RTX PRO 5500 is a workstation chip, less powerful than the data-center processors that are barred from China, and it is export-compliant. Reuters also reported that Nvidia has indicated it could build no more than three million of the cards globally in 2027, and that total orders from Chinese customers exceed six million. Secondary reports citing the Reuters story say Nvidia has told customers fulfillment might reach only 5% to 10% of requested quantities.
Why it matters: The $35 billion figure measures intent, not delivered hardware. What it does show is how tight computing supply has become in China and how export-compliant products are absorbing demand that restricted chips cannot meet. Washington's curbs have pushed Chinese firms toward domestic suppliers, and Beijing encourages homegrown chips, so the order is a test of how far a lower-spec Nvidia product can fill the gap. For related context on Nvidia's shrinking China share, see AI Watch. Connect to: AI Intelligence · Semiconductor Intelligence · Hardware Watch
OVHcloud Prepares Its First AI Model as 110 MW of Capacity Sits Open
What happened (October 20, CEO statements): After publishing annual results, CEO Octave Klaba told Reuters that OVHcloud has 250 MW of capacity connected, 140 MW of it in operation, leaving 110 MW ready for new clients, and that its first AI model should launch in early 2027. The model is reported to use a mixture-of-experts design. OVHcloud also reported its first positive levered free cash flow since its 2021 IPO and is targeting revenue growth above 10% in 2027.
Timeline conflict: EconoTimes notes that La Tribune reported an initial open-source language model in late November 2026, while OVHcloud's published results give no model date. It is unclear whether these are separate milestones. OVHcloud plans to detail its AI strategy at an AI Summit on November 19.
Why it matters: Spare, already-powered capacity is an asset in a market where grid access is the constraint (see item 6). But capacity only becomes revenue when customers show up. Klaba has also been sceptical of national AI-gigafactory strategies for lacking a customer base to match their capacity. OVHcloud's position, powered capacity first and demand second, is the opposite bet to projects that are still waiting for power. Connect to: Cloud Computing · Enterprise AI Intelligence
Elice's Weak Kosdaq Debut Tests AI Infrastructure Valuations
What happened (October 20, confirmed): Elice, a South Korean AI cloud and software provider, priced its IPO at the top of its range at 90,500 won and raised about 201.1 billion won (roughly $151 million), per TechNode. The stock closed about 32% below the IPO price after falling as much as 37.6% intraday, according to one account. The institutional book was 388.84 times subscribed, per Seoul Economic Daily, and Dealroom reports first-half revenue up 200% and a swing to operating profit. It was the largest Kosdaq IPO of 2026.
Context: Reports link the debut to continued selling of semiconductor stocks in Seoul and note that Nvidia-backed Firmus abandoned a $5 billion Australian IPO less than two weeks ago.
Why it matters: Oversubscription did not guarantee aftermarket support. Public investors are pricing AI infrastructure on revenue visibility and financing costs, not on demand headlines. One debut does not prove the business is failing; the test is whether new capacity converts into lasting cash returns. Connect to: VC Intelligence · IPO Watch · AI Infrastructure Watch
AI and Enterprise Technology
Microsoft Introduces Controls That Treat AI Agents as Insider Risks
What happened (October 19, company announcement): Microsoft announced new capabilities in Microsoft 365 E7 and Project Perception, per SiliconANGLE and Help Net Security. Dynamic Agent Risk scores each agent continuously, Dynamic Conditional Access extends Entra controls to agents, and Guardian Agents supervise other agents' actions. Microsoft is also expanding discovery of locally running agents by 35 types, for a total of 60, including Claude Code, OpenAI Codex and GitHub Copilot. Microsoft did not give availability dates or pricing for the new controls.
Why it matters: The announcement frames enterprise agents as insiders to be monitored and contained, and it covers third-party agents as well as Microsoft's own. That is a governance product aimed at the same problem enterprises raise about agent permissions. As one analysis notes, these are announced controls, not independent evidence that they work. Connect to: Enterprise AI Intelligence · Cybersecurity Watch · Microsoft Company Analysis
Semiconductors and Infrastructure
Nvidia Pushes Beyond GPUs as Modal Adopts Vera CPUs for AI Agents
What happened (October 19, customer announcement): Modal, an AI cloud startup, said it will run Nvidia's new Vera CPU to host AI agents alongside its Intel and AMD chips, per reports citing Reuters. Modal said early tests showed Vera completing up to twice as much work as comparable x86 hardware under heavy concurrent load. Those are Modal's own results, and the move required significant software changes. Nvidia has projected at least $20 billion in Vera sales this fiscal year and puts the CPU market at about $200 billion. Modal recently raised $750 million at a $15.75 billion valuation, and OpenAI, Anthropic, Oracle, and Perplexity have said they plan to use Vera.
Why it matters: Agents spend time on the CPU between model calls, and Modal argues that CPU time saved is GPU time saved. If agent workloads favor Arm-side GPU affinity, Nvidia gains a way into a socket long held by Intel and AMD. Modal is mixing all three suppliers, so this is a first design win, not a displacement. Connect to: Hardware Watch · Semiconductor Intelligence · Enterprise AI Intelligence
Goldman Warns Power Delays Could Keep Half of Planned Capacity Off Schedule
What happened (October 20, analyst estimate): Axios reported an exclusive on a Goldman Sachs report expecting global data center power demand to rise 170% between 2025 and 2030, with more than 60% of that growth coming from the US. Goldman estimates about $7.6 trillion of capital spending from 2026 to 2031 across computing, data centers and power, and notes roughly 2,600 GW of proposed generation and storage projects waiting for grid connection, though many may never be built. A separate summary of the report says about half of planned capacity may arrive on schedule over the next couple of years, measured by capacity rather than facilities. A second report says that on October 19, FERC issued directives to six regional grid operators to justify tariffs governing how large loads such as data centers connect. Verify that FERC action against FERC's own notice.
Why it matters: Goldman's framing is that the AI race is increasingly a race for power. For enterprise buyers, delivery dates for capacity depend on utilities, not only chip supply. Tech companies are reportedly considering funding generation and transmission directly, which would shift cost and risk onto their balance sheets. Connect to: AI Infrastructure Watch · AI Infrastructure Special Report · Cloud Computing
Cybersecurity
Critical BeyondTrust Remote Support Flaw Was Exploited Before Disclosure
What happened (disclosed October 15, reported through October 16; earlier story still active): BeyondTrust disclosed a critical vulnerability in Remote Support, CVSS 9.3, tracked in advisory BT26-05, per Infosecurity Magazine. Google Threat Intelligence and Mandiant attribute exploitation since August 29 to a suspected China-linked group tracked as Armory Castle, which installed a custom webshell and tried to extract credentials. The vendor says versions through 26.2.3 are affected and that it withheld technical details while customers update. Google warns of potential widespread exploitation now that the flaw is public.
Why it matters: Remote support appliances connect to many endpoints and may hold the credentials to administer them. Installing the patch closes the hole but does not undo activity that may already have occurred, so teams should check for earlier compromise, and the same product line has been targeted by state actors before. The attribution to a China-linked group is Google's and BeyondTrust's assessment based on operational methods.
Connect to: Cybersecurity Watch
Startups, Funding and M&A
Salesforce Agrees to Buy Braze for About $5.66 Billion
What happened (October 20, definitive agreement): Salesforce signed a definitive agreement to acquire customer engagement platform Braze for about $5.66 billion in enterprise value, paying $47.50 in cash per Braze Class A share, per Salesforce's 8-K press release and CNBC. Braze closed at $30.04 on Monday and finished Tuesday up about 51% at $45.57, while Salesforce shares fell 2.6% in premarket trading. The deal is expected to close in Salesforce's fiscal 2028, subject to Braze shareholder and regulatory approvals, and Salesforce said it does not change fiscal 2027 guidance. Braze serves about 2,700 customers and will keep its brand, with CEO Bill Magnuson staying on.
Why it matters: Salesforce is adding real-time engagement and AI decisioning to Marketing Cloud, tying it to Data 360, Agentforce, and its wider platform. The cash price carries a large premium, and the long timeline to close leaves regulatory and shareholder risk open. Investors reacted cautiously on the Salesforce side.
Connect to: Tech M&A · Company Intelligence · The Term Sheet
Why It Matters
Demand is real, but it is hitting three bottlenecks. Chinese buyers want more export-compliant Nvidia hardware than can be built. European and Asian cloud providers are discovering that powered capacity is scarce, but customers and public-market patience are not unlimited. And the power grid sets the pace for everything else.
The market is repricing proof over promise. Elice's debut and the abandoned Firmus IPO suggest investors want revenue visibility. OVHcloud's cash-flow milestone fits the same preference for demonstrated returns.
Agents are becoming a governance and architecture problem. Microsoft is selling containment for insider-style agents, and Modal's test results suggest agent workloads may reshape server CPU choices. Both point to production agents needing different infrastructure and controls from chatbots.
What to Watch Next
- NVIDIA GTC Berlin, today. Jensen Huang's keynote is at the Tempodrom on October 21, with sessions on agentic AI, robotics, open models and physical AI. Watch for any comments on China demand, Vera, and European sovereign-AI partnerships.
- Earnings after the close today. Meta, Tesla, Lam Research, SAP, Texas Instruments and IBM report. Listen for capex guidance, power and financing commentary, and enterprise AI demand.
- Nvidia and the RTX PRO 5500. Whether Nvidia, Chinese buyers, or Washington comment on the reported orders, and how much is actually delivered.
- OVHcloud's AI Summit, November 19. The company plans to detail its AI strategy, which should clarify the model timeline.
- FERC and grid-operator responses. Confirm the October 19 directives and how operators justify their large-load tariffs.
- Braze approvals. Shareholder vote, regulatory review, and any competing interest ahead of the fiscal 2028 target close.
- BeyondTrust patching. Signs of broader exploitation and any further advisories.
The CODEW Stat
388.84x vs. -32% — Elice's institutional book was roughly 389 times covered, yet the stock closed about 32% below its IPO price on day one. The gap is the story: heavy IPO demand did not translate into aftermarket support for an AI infrastructure name. Pair that with a reported $35 billion order that cannot be filled and a grid that cannot connect new load fast enough, and the signal is the same across all three. Demand is not the scarce input; credible supply, power, and proof of returns are.
Sources: Reuters, CNBC, Axios, Salesforce's SEC filing, Microsoft announcements as reported by SiliconANGLE and Help Net Security, Infosecurity Magazine, Google Threat Intelligence and Mandiant as reported, TechNode, Seoul Economic Daily, Dealroom and secondary summaries, covering October 15 to 21, 2026. Where a story is relayed through a secondary outlet, the item says so.
Confirmed announcements, media reports based on anonymous sources (Nvidia orders), analyst estimates (Goldman), and pending transactions (Braze) are labeled separately. Figures, prices and dates should be verified against company announcements before publication.
Reviewed by Erwin Castro
on
Wednesday, October 21, 2026
Rating:
No comments: