CVCOVANewsComputing, Opportunities, Value & Analysis
Tech News5 min read

Pricier AI chips and an Nvidia in crisis: what's at stake for Latin American startups

Renting one of the most widely used GPUs for artificial intelligence now costs about 50% more than it did a year ago, according to an industry index. Now a Wall Street Journal report on how Nvidia financed its own customers raises further questions about how those prices are set.

By COVA News newsroom
black and gray audio mixer

Photo: Dimitris C on Unsplash

A year ago, renting an Nvidia H100 for one hour cost about $2. It is the chip used to train and run a large share of artificial intelligence models. On November 9, 2026, the index from the firm Silicon Data stood at $3.01. For a Latin American startup that has no servers of its own and depends on the cloud, meaning it rents remote computers from companies such as Amazon or Google, that increase hits its budget directly. This week another piece of the puzzle emerged: a Wall Street Journal report describing how Nvidia helped finance some of the customers that were buying those chips.

The report was published between November 12 and 13. Its contents are known from what other outlets have relayed, because the original is behind a paywall and could not be consulted directly. It is neither a lawsuit nor a regulator's finding. According to those summaries, then-CEO Jensen Huang pushed for the company to finance startups and small cloud providers so they could acquire its GPUs, the processing cards that perform AI calculations. Chief financial officer Colette Kress reportedly warned that some of those companies could not afford them and put the brakes on several commitments. Among them was reportedly a proposed deal with OpenAI worth about $250 billion, which Kress allegedly cut in half. That figure is not backed by public documents. A revised version of the deal was signed on August 17, and at the end of that month Nvidia paused some financing and added reviews by its executives and its board of directors, again according to the reporting.

Manufactured shortage or surplus chips?

The story reached the front page of Hacker News under an eye-catching headline about a "manufactured shortage." That phrase does not come from the Wall Street Journal: it comes from the site TechPowerUp. According to summaries of the newspaper's report, the picture it paints is almost the opposite: an oversupply. At an August meeting, Huang reportedly said that the silicon wafers and memory Nvidia had committed to buying were arriving "like an avalanche," because the major cloud providers were starting to use chips they designed themselves. Along the same lines, Wccftech reported that Nvidia had allegedly asked TSMC, the manufacturer that produces its chips, to reduce future orders. The idea of a deliberately engineered shortage is circulating in media outlets and forums, but the site Explainx describes it as unverified: no regulator has reached that conclusion, and Nvidia has not acknowledged any such strategy.

Behind all this lies an older debate over so-called "circular financing," in which a supplier invests in its customers and those customers then buy from it. Nvidia has invested in OpenAI and in CoreWeave, a cloud provider specializing in AI. Bloomberg also names xAI (now part of SpaceX), Mistral, Nscale and Nebius among the companies with deals of this kind. Huang rejected that interpretation at a Goldman Sachs conference: he argued that the amounts are small relative to the size of the business and that each investment stands on its own merits.

So why are prices rising? The analysts cited in the coverage do not blame any maneuver by Nvidia. SemiAnalysis found that the price of renting an H100 for one year rose from $1.70 to $2.35 per hour between October 2025 and March 2026, nearly 40% more, and attributes this to demand from coding agents and reasoning models, which consume a great deal of compute. The firm Aquanode adds two manufacturing bottlenecks: CoWoS packaging, the technique that joins the chip to its memory, and the high-speed memory itself, known as HBM.

The impact on the region, still unmeasured

For Latin America, there is no direct data on how all this is moving cloud prices. The most concrete sign is physical: El Financiero reported on November 11 that in Mexico the most sophisticated servers take between 27 and 30 weeks to be delivered, and the simplest ones between 16 and 18 weeks, due to shortages of SSDs, memory and GPUs. As for the cloud, the available reference comes from the blog of Spheron, a GPU provider with a commercial interest in the subject. According to that blog, AWS in São Paulo is the only region in the area with H100s available locally, Google Cloud in that city offers only the older T4 model, and a general-purpose virtual machine there costs about 1.6 times as much as in Iowa.

All of this comes at the worst possible moment for the company. On November 12, Nvidia announced Huang's resignation as CEO and named Kress interim chief executive. CNBC and TechCrunch link his departure to a personal scandal involving an alleged Chinese spy, a separate matter from the financing report, which came out afterward. The stock closed below $4 trillion in market capitalization for the first time since July 2025, and Motley Fool puts it 32% below its peak.

The next key date is November 17, when Nvidia is scheduled to report earnings and could address the report. For companies in the region, questions remain open: whether the oversupply described by the Wall Street Journal will end up being reflected in what cloud providers charge, when more cutting-edge chips will reach Latin American data centers, and how much real access local startups have to compute credits. None of the available sources measures this yet.

Related news

Tech News

Bitcoin falls below $60,000: what changes for people who save or send remittances with crypto

Bitcoin broke below $60,000 for the first time since July, after record outflows from a BlackRock fund and a sell-off in AI-linked stocks. In Latin America, the blow is not the same for people who save in bitcoin as for those who send money with digital dollars.

Tech News

Wall Street enters correction territory and Anthropic falls below its first close after the largest IPO in history

A day after Anthropic's record debut, the S&P 500 and the Dow entered correction territory and the Nasdaq erased its gains for the year. The sell-off reopens the debate over an artificial intelligence bubble and was also felt on the Mexican Stock Exchange.

Tech News

More than 100,000 signatures call for Jensen Huang's return to Nvidia as the Megaspeed case grows

A petition backed by Elon Musk and thousands of purported employees seeks to put Huang back in charge of Nvidia. The same day, The Wall Street Journal revealed how Megaspeed allegedly channeled nearly $12 billion worth of chips to Chinese companies.