Nvidia et TSMC accélèrent la production des puces H200, alimentées par la demande chinoise

La course à la puissance de calcul s'intensifie. Nvidia pousse sur l'accélérateur pour ses puces H200, et c'est avec son partenaire de longue date, TSMC, que le géant du GPU prépare sa prochaine offensive.

Une alliance stratégique face à la demande

Alors que les restrictions sur les exportations de technologies avancées vers la Chine créent un paysage complexe, Nvidia adapte sa stratégie. L'objectif est clair : répondre à une demande qui ne faiblit pas, notamment de la part d'acteurs chinois cherchant à sécuriser leur approvisionnement en puces haute performance pour l'IA et le calcul intensif.

TSMC, le partenaire incontournable

La production de ces semi-conducteurs de pointe repose sur les épaules de TSMC, le fondeur taïwanais qui domine le marché des procédés les plus avancés. Cette collaboration renforcée souligne à quel point la chaîne d'approvisionnement mondiale reste fragile et hautement spécialisée. Une simple perturbation peut envoyer des ondes de choc à travers toute l'industrie tech.

Le timing est crucial. Alors que la concurrence dans l'IA générative fait rage, la possession des meilleurs GPU devient un avantage stratégique décisif. Nvidia le sait et positionne ses pions en conséquence. Une manœuvre qui, ironiquement, pourrait bien financer la prochaine vague d'innovation... ou simplement gonfler encore un peu plus la valorisation déjà astronomique de l'entreprise. Après tout, en finance, on appelle parfois "investissement dans l'avenir" ce qui ressemble surtout à une course effrénée pour justifier des multiples de valorisation toujours plus vertigineux.

Chinese firms order millions of chips as Nvidia balances supply gaps

The talks with TSMC, as well as the size and price of Chinese orders, have not been reported before. Nvidia has priced the H200 chips at around $27,000 each, depending on the buyer and order size.

Two sources said the company will offer two chip variants to Chinese customers: the standalone H200 and the GH200 Grace Hopper superchip, which combines the Grace CPU with the Hopper GPU.

Of the 700,000 units Nvidia currently has, about 100,000 are GH200s, with the rest being H200. The first deliveries will come from this existing inventory and are scheduled to reach clients before the Lunar New Year holiday in mid-February. Additional supply will follow once TSMC ramps up.

Chinese firms view the H200 as a serious step up from what they can currently access. The now-blocked H20, a weaker chip made specifically for the Chinese market, is no longer available after Beijing banned it. But the H200 delivers roughly six times the performance, according to people familiar with the matter.

The eight-chip module is priced at about 1.5 million yuan, more than the H20’s 1.2 million yuan price tag. Even so, it’s still cheaper than gray-market options, which are going for over 1.75 million yuan.

ByteDance is already gearing up to spend 100 billion yuan on Nvidia’s chips in 2026, up from 85 billion yuan in 2025, if Chinese regulators approve the H200 imports.

Beijing still undecided on greenlighting incoming chip shipments

Despite the U.S. now allowing H200 exports, Chinese regulators haven’t given the all-clear. They’re worried that letting in more advanced foreign chips could slow progress in China’s own semiconductor sector. Officials haven’t blocked the shipments, but they haven’t signed off either.

Local chipmakers have built products that match the H20, but nothing yet rivals the H200. One idea that’s reportedly being discussed in Beijing is to tie every imported H200 chip to a mandatory purchase of a set amount of locally made chips.

This plan would let China’s domestic players stay in the game while letting internet giants like ByteDance keep scaling up.

Nvidia responded to a request for comment saying it manages its supply chain actively. A company spokesperson added, “Licensed sales of the H200 to authorised customers in China will have no impact on our ability to supply customers in the United States.”

The spokesperson also reportedly said, “China is a highly competitive market with rapidly growing local chip suppliers. Blocking all U.S. exports undercut our national and economic security and only benefited foreign competition.”

The H200 is part of Nvidia’s Hopper architecture and is made using TSMC’s 4-nanometer process. Even though Nvidia is also working on newer chips like Blackwell and the upcoming Rubin, this sudden surge in Chinese demand is pushing it to expand H200 output fast.

Sources said Nvidia hasn’t finalized exactly how many more chips it’ll ask TSMC to build, but the target is to keep up with massive demand while avoiding deeper supply shortages in other regions.

That balance just got a lot trickier.

Join Bybit now and claim a $50 bonus in minutes