Tech Newsround – Nvidia, OpenAI & CXMT

Nvidia & OpenAI – Cost leadership.

• It looks like Nvidia is going to provide a guarantee of $500bn that will allow OpenAI to lease the 10GW centre in Ohio, which is crucial because any wobbles could allow suppliers to renegotiate prices upwards.
• The site in question is the 10GW, $500bn site in Ohio, which broke ground on the first 800MW in March, which I would expect to come online in 2028.  
• I think that this is crucial to OpenAI because there are signs that its move to secure capacity early has resulted in it signing contracts at relatively low prices, which could give it a big edge over its arch-rival Anthropic in 2027.
• Up until 9 months ago, the industry standard was $10bn per GW, but the parabolic growth in revenue at Anthropic has left it very short of capacity (like everyone else), which in turn has led to a very rapid increase in compute capacity pricing.
• This has resulted in Anthropic and Google signing deals for capacity that are as much as 5x the previous price, meaning that they too will need to increase price substantially.
• OpenAI, on the other hand, secured most of its capacity some time ago, which leads me to think that its pricing is far superior, meaning that it should be able to undercut everyone else in late 2026 and 2027.
• Hence, if I were OpenAI’s supplier, I would be looking for any excuse to renegotiate the price upwards and the inability to guarantee a lease contract would be just such an opportunity.
• This is why Nvidia stepping in is significant, as it removes this risk and should allow OpenAI to continue to benefit from its superior cost structure.
• There will be a lot of discussion around circular deal-making, and this is certainly an issue, but if OpenAI can become a cost leader by a wide margin, this makes its outlook for 2027 and an IPO much better.

CXMT – IPO Bounce.

• CXMT has rallied nearly 500% in its debut to become the largest Chinese listed company as the AI-driven frenzy in memory spreads to China, even though it does not make the advanced memory used in AI systems.
• At RMB49.5, CXMT has a market cap of $539bn and is trading on a PER of 1,800x, which makes even SpaceX look like a bargain.
• However, CXMT is growing fast and is expected to reach Micron’s capacity at the end of this year and 500,000 wafers by the end of 2028.
• CXMT is doing well largely on the back of the global DRAM shortage that is plaguing the industry and has been caused by the AI rollout.
• For the developed market memory players, HBM memory has higher margins, and so they have switched capacity from commodity DRAM for consumer electronics devices to HBM for the data centre.
• This is the gap that CXMT is filling, as being located in China, it cannot make advanced semiconductors needed for AI data centres as it is unable to purchase the requisite equipment.
• Hence, when the supply for DRAM normalises (2028 according to Micron), then CXMT is likely to see a big correction, but between then and now, it looks like good times are ahead.
• Many Chinese IPOs have had very strong opening days and then collapsed, and so I am far from convinced that CXMT will hold these levels even with the current levels of demand that it has.