Nvidia's $500 Billion Compute-Financing Play: The Odds and Risks

---

TL;DR

• Nvidia is teaming up with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR on a $500 billion financing plan intended to turn compute into an “investable asset class.” CEO Jensen Huang is the same person who said a year ago that “once Blackwell starts shipping, Hopper won’t be wanted even for free,” while BlackRock CEO Larry Fink compares the plan to his own experience getting involved in the mortgage-backed securities (MBS) market in the 1970s.
• Rising rental prices for older chips back real demand, but risks on the demand side — data center oversupply and Chinese open-source models achieving strong performance with less compute — are also flagged within the same article.
• Also covers NTT’s compact domestic LLM “tsuzumi 2,” OpenAI’s expanding partner network in Japan, a Kobe University study measuring AI’s persuasiveness in moral judgment, and the extended expansion of Claude Code’s weekly usage limits.

Top Story: Nvidia’s $500 Billion Compute-Financing Play — The Odds and Risks

Nvidia is working with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR on $500 billion in financing intended to turn compute into an investable asset class. CEO Jensen Huang told CNBC, “This is really the first example of tech chips becoming an investable asset class. These are revenue-generating assets — long-lived, fungible, and flexible” (The Verge). BlackRock CEO Larry Fink, also speaking to CNBC, said, “This is just the beginning. I see this the same way I saw the mortgage-backed securities (MBS) market when I first got involved in the 1970s — as the next frontier of financial engineering.”

Technical Take

This claim is in tension with Huang’s own past remarks. A year ago, ahead of volume shipments of the next-generation Blackwell GPU, Huang said that “once Blackwell starts shipping, Hopper won’t be wanted even for free” and that “there are use cases where Hopper fits, but not many.” The prospect that previous-generation chips can lose value during generational transitions calls for an explanation of the conditions under which they remain “long-lived” assets. On the other hand, rising rental prices for older chips are backed by measured data: Silicon Data projects the increase will continue through 2028, and one cloud provider nearly doubled its Blackwell B200 rental price for one customer upon contract renewal. Evidence that “demand is real” and the claim that “chips as an asset retain value over the long term” are separate propositions, and the latter does not follow from the former.

Business Take

The organizations involved include major asset managers, investment firms, and financial institutions. The fact that Fink frames a plan to treat GPU compute as a revenue-generating financial asset by comparing it to the 1970s MBS market says a lot about the nature of this financing. On the other hand, as former hedge fund manager Mark Rubinstein has pointed out, the MBS market collapsed because of an oversupply of mortgages. The article also flags two factors in the AI industry that could threaten the premise of ever-growing chip demand: data center supply saturation, and Chinese open-source models achieving reasonably strong performance with less compute. That said, these are qualified observations, and there is not yet any confirmation that oversupply has actually shown up in prices.

Contrarian Take / What’s Overlooked

On the surface, the financialization of compute itself looks like confirmation of the thesis that “value accrues to infrastructure/distribution.” But what supports this financing plan is the unverified premise that chip demand will keep growing. The health of the financing hinges on whether compute cash flows (chip rental demand) play out as expected. Because Huang’s current description of chips as “long-lived” sits uneasily with his earlier remarks about previous-generation Hopper GPUs, the credibility of that premise should be discounted accordingly.

Implications and Positioning

The direct impact on smaller businesses is limited, but this matters as a signal for the medium-term outlook on AI API and compute costs. If the plan stabilizes long-term funding supply, GPU capacity could expand, putting downward pressure on API and rental prices in the medium term. Conversely, if the capital inflow doesn’t match real demand, the risk of a supply shock or a sudden price swing would keep building up. Baking “AI compute costs will keep falling” into unit-economics calculations as a fixed parameter is risky, and it’s worth reviewing this kind of change in the supply-side financial structure roughly once a year.

Signal/noise verdict: The claim that “compute is an asset class” is noise. Huang’s current description of chips as “long-lived” is in tension with his remarks a year ago, and the risks are laid out within the same article, so this should not be taken at face value — it reads as a narrative meant to smooth fundraising. On the other hand, the underlying measured data on rising rental prices is signal and worth watching. Confidence level: moderate (both whether the financing plan goes through and how prices move can be verified over the coming quarters).

Other Key Topics

NTT’s Compact Domestic LLM “tsuzumi 2” Bets on Japanese-Language Cost Efficiency

NTT has developed “tsuzumi 2” as a 28-billion-parameter dense model (a design that uses all parameters during inference), positioning it as a “sovereign AI” that doesn’t depend on other companies’ LLMs. Its proprietary tokenizer incorporates insights from morphological analysis, designed to increase the number of Japanese characters per token and thereby reduce token consumption. It was pretrained on a corpus of roughly 10 trillion tokens and scored highly on the Japanese instruction-following benchmark “M-ifeval”; the article also cites cases where it followed complex instructions, such as list-formatting requirements, more faithfully than models from other companies (ITmedia AI+). Technically, the choice of scale — 28B parameters, runnable on a single GPU — is a design worked backward from ease of deployment in real operations rather than maximizing performance. From a business standpoint, the proprietary tokenizer and internally managed training process are potential upstream differentiators. Rather than competing with frontier models on general chat performance, this reads as a niche strategy aimed at industries where on-premises operation and data sovereignty are mandatory. Given the potential to reduce token consumption for Japanese-language processing, it’s worth considering as a diversification option away from relying solely on general-purpose API models.

OpenAI Expands Partner Network in Japan, Targeting the SME Long Tail

OpenAI announced it is investing $150 million globally in the “OpenAI Partner Network” and aims to train 300,000 certified consultants by the end of 2026. In Japan, alongside major system integrators like Fujitsu and NTT Data, three new companies — Leave a Nest Knowledge, Simplex, and Recursive — took part in a roundtable discussion. Simplex said that letting AI draft design documents and requirements specifications has enabled even junior engineers to produce deliverables of consistent quality, while Recursive explained that it replaced a CRM SaaS product with an in-house tool built using Codex (ITmedia AI+). There’s little technical novelty here; the theme is reducing adoption friction. From a business perspective, it’s telling that OpenAI is using the partner network to supply industry knowledge, change management, and systems integration that it cannot provide alone. This expands a network previously centered on major system integrators toward regional and smaller businesses, supporting the view that near-term revenue opportunity lies in implementation support and industry expertise rather than in the model itself. AI implementation support for SMEs could gain some defensibility if it moves beyond a pure consulting model toward industry-specific templates.

Kobe University: ChatGPT’s Counterarguments Reverse Over 30% of Moral Judgments — Older Adults Especially Persuadable

A Kobe University research team had ChatGPT argue against responses given by 56 participants aged 18-30 and 74 participants aged 65 and older to two versions of the trolley problem (the switch scenario and the footbridge scenario), then examined whether the participants’ judgments changed. The result: 32.31% of participants reversed their judgment on the switch scenario and 36.92% on the footbridge scenario. Prior research found that people incorporated advice from others into their judgments roughly 20-30% of the time, so both figures in this study exceeded the top of that range. The study also found that older adults with reduced cognitive function were more susceptible to persuasion. The findings were published on August 3 in Computers in Human Behavior (ITmedia AI+). This research measures human susceptibility to persuasion rather than model performance, and the fact that AI counterarguments exceeded the range reported in prior interpersonal-advice research gives decision-support products a concrete persuasion risk to account for. For consumer products involving decision support (healthcare, financial advice, etc.), persuasive power needs to be treated as a risk rather than a feature. In designing AI chat products aimed at older adults or other decision-vulnerable users, design choices that dampen persuasiveness — such as presenting options rather than using an argumentative, steering UI — may become necessary.

Claude Code’s Expanded Weekly Limits Extended Again Through August 31, With Capacity Strain Acknowledged

Anthropic has again extended, through August 31, its campaign offering a 50% increase to Claude Code’s weekly usage limits (covering Pro, Max, Team, and Enterprise plans). The company explained, “We’d like to make this change permanent, but demand for AI models is high, and capacity may be tight over the coming weeks.” The campaign began on May 13, and this marks one of several extensions. The article also reports that as background, active users of competitor OpenAI’s “Codex” have surpassed 15 million (ITmedia AI+). The fact that Anthropic itself is using the phrase “capacity strain” can be read as aligning with the real-demand signal of rising chip rental prices on Nvidia’s side discussed in the top story. The pattern of repeated extensions without committing to permanence looks like a user-acquisition tactic aimed at competing with rivals, but it can also be read as a sign that the current relaxed terms don’t currently pencil out against cost. If you’re building Claude Code into a production workflow, treat the current expanded limits as a time-limited promotional condition and plan costs assuming a return to normal limits after September.

Try This Week / Hype to Ignore

Try this week: While Claude Code’s expanded limits are still in effect, it’s worth front-loading batch-style usage you’d normally hold back (large-scale refactors, exhaustive test generation, etc.) to prepare for the limits shrinking back after September.

Hype to ignore: The phrase “compute is an asset class” itself. Huang’s current description of chips as “long-lived” is in tension with his remarks a year ago, and oversupply risk is flagged within the same reporting, so this should be treated as a narrative designed to smooth fundraising — it’s premature to take it at face value as “the birth of a new, stable asset class.”