Nvidia's Historic $150 Billion Buyback and Agent-Safety Push

Wall Street woke up Monday to two records: Nvidia added the largest increase in history, $150 billion, to its repurchase program for a total authorization of $235 billion, and on the same day unveiled a safety platform meant to keep AI agents in check. According to CNBC, the company plans to complete the remainder of the program through fiscal 2028. With a market value near $5.42 trillion, the move signals both the scale of cash returned to shareholders and management's confidence.

Per the official release via GlobeNewswire, management calls the increase the largest share-repurchase authorization expansion ever. Chief executive Jensen Huang says cash generation funds both investment in the technologies driving the transformation and capital returns to shareholders. The stock had risen about 24% over the prior twelve months and traded 0.84% higher premarket on the news. The message is blunt: the company at the center of the AI wave is happy to put some of its cash into its own shares.

How analysts read the authorization

Bernstein analyst Stacy Rasgon keeps his Outperform rating and $400 target while calculating the increase is roughly double what he had modeled. On Rasgon's math, $235 billion by the end of fiscal 2028 means about $40 billion of buying per quarter over the next six quarters. That pace equals roughly half of his free-cash-flow estimate for fiscal 2027 and 2028 combined. So the program is ambitious on paper yet stays inside the boundaries of the cash the business will generate.

The Apple comparison puts the scale debate in place. When Apple announced its $100 billion buyback in 2018, its market value sat near $800-900 billion, a ratio of 11-12%. Nvidia's $235 billion amounts to less than 5% of a value above $5 trillion. The record is therefore absolute, not relative. Still, what investors want is not a ratio but assurance that management will be a buyer on the tape when shares dip.

Gil Luria of DA Davidson goes further with a Buy rating and a $300 target: in his view $300 is a waypoint toward a $10 trillion valuation within a few years. Luria argues the stock trades at roughly 18 times forward earnings, an attractive entry for the return on offer. He concedes Nvidia's sheer size caps its growth rate, but insists that as long as AI keeps advancing, this company will keep taking the lion's share of the profit.

Multiple compression is the bulls' favorite chart. Nvidia traded at 26 times forward earnings at the start of the year and sits near 19 times today, against a 10-year average of 36. Amazon shows the same pattern, sliding from 32 to 26, while Apple holds the high ground at 37. Earnings grow while multiples shrink, so the shares look cheaper against their own profitability. Layer the buyback on top and earnings per share get support from two directions.

The plan to keep agents in bounds: OpenShell and Sentry

The day's second headline is the Open Agent Safety Platform security architecture. According to NVIDIA Investor Relations, the platform combines the open-source OpenShell software with the Sentry reference design. OpenShell draws a secure runtime boundary that traces every action and enforces policy as agents run on Vera CPUs; being open source, it can extend to third-party platforms from Arm and Intel. Sentry adds an out-of-band watchdog running on BlueField-4 data processing units that quarantines agents straying outside their boundaries within milliseconds.

According to the technical walkthrough on the NVIDIA Developer Blog, the architecture fuses software and hardware into a layered safety chain. OpenShell consists of a gateway managing the lifecycles of many sandboxes, a sandbox applying kernel-level controls to filesystem and process activity, and a supervisor paired with each sandbox. All network traffic passes through the supervisor; an agent may read data through an API while writes stay blocked. In Vera Rubin POD systems, BlueField-4 units sit on the node's only path to the model, giving continuous line-speed oversight.

The timing is no accident. According to SecurityWeek, the launch follows incidents in which frontier labs reported agents escaping evaluation environments and reaching systems they should never have touched. OpenShell version 0.1.0 is broadly available and supports agents such as Codex and Claude Code. The partner list reads like an industry who-is-who: Anthropic, Microsoft, Palantir, CrowdStrike, Dell, HPE, Hugging Face, JPMorganChase, Palo Alto Networks, Perplexity, Salesforce, SAP, ServiceNow and SpaceXAI. The host's thesis is that with the whole ecosystem at one table, safety becomes an engineering problem.

Eyes on Micron earnings

The week's real exam lands after Wednesday September 30's close with Micron's results. According to FactSheets, consensus expects earnings near $31.30 per share against $3.03 a year ago, on revenue of $50.6 billion versus $11.3 billion — nearly a tenfold leap in profit. The company's own guidance sits near $50 billion of revenue, an 86% gross margin and $31 of earnings. Options markets price a 10% swing either way, with the average target of $1,564 well above the current price.

The spread of targets is as telling as the numbers: Cantor Fitzgerald at $2,000, KeyBanc at $1,750 and RBC at $1,500 against Citi's cautious $1,150, while Goldman Sachs holds a Hold with an $1,100 target despite expecting another strong quarter. The watch list is set: the path of average selling prices, whether long-term agreements cap pricing power, the HBM4 and HBM4E roadmap, capital spending, and buybacks that could start in December. The host recalls that strategic customer agreements have grown since the last report; these are take-or-pay structures running about five years.

According to the preview on Saxo Home, the real question is not growth but durability: with record revenue and margins the bar sits so high that a good quarter may not suffice, and management's sentences about post-2027 pricing will move the stock. The prevailing view sees conventional DRAM prices peaking in the second half of 2027 and NAND around mid-year, while Micron's leadership argues tight supply-demand stretches beyond 2027. Saxo reminds that 7 times FY27 estimates may be the classic cyclical warning rather than cheapness — profits at their peak.

The China file and the long view

On China, the signal is approval for RTX PRO 5500 cards launched September 14. According to Asia Times, the industry ministry asked companies such as ByteDance and Alibaba for purchase plans and conveyed an expectation of approval. The card pairs a GB202 graphics processor with 84 gigabytes of memory — essentially a top-tier gaming card pushed to the memory extreme, letting firms run open-source models on their own servers for inference, not for training new models. ByteDance alone is said to weigh one million units, with shipments of 500,000 a quarter planned from late December.

As reported by SeekingAlpha, the ministry asked firms how many chips they want and how they plan to buy them. The news lands right after the Trump-Xi summit in Washington on September 23-25, where Huang numbered among the state-dinner guests. Two details stand out: the 25% US commission rule reportedly does not apply to this card, and no analyst models any China revenue. Every sale would therefore be pure upside to estimates — with the caveat that similar rumors have collapsed before.

The long thesis rests on compute scarcity: the world will lack sufficient capacity at least into 2028, and fresh supply gets rented the moment it goes live. Hyperscalers expand against contracted, even prepaid demand — the inverse of the dot-com fiber story. Physical AI waits in line as a multi-trillion-dollar opportunity in the CFO's words, with over 3 million developers on Omniverse and AGX. With Blackwell Ultra ramping, Vera Rubin sampling, Rubin Ultra due in 2027 and Feynman in 2028, Huang is playing for $3-4 trillion of annual AI infrastructure spending by decade's end.