Computex 2026 Day 1 – GTC Taiwan or Computex?

Jensen sets the agenda for Computex 2026

• Nvidia headlined another Computex, making the case for AI, rounding out its edge device portfolio and updating its software and tools so that nobody strays from the Nvidia ecosystem.
  • First, AI doomsayer debunking: where Jensen took the “AI kills software” and “AI takes everyone’s job” concerns and blew them up with data as opposed to anecdotes.
  • Jensen’s opinion (with which I agree) is that AI can write software, but it needs supervision and orchestration, which is why humans are in greater demand than ever.
  • For example, commits (uploads) to GitHub are up 3x in 2026 alone, and job postings for US software engineers are up 15% YoY and are at an all-time high.
  • This indicates that AI is making software engineers more productive, which in turn means that rational employers will want more of them, not less.
  • It also points to the SaaSpocalypse for the Enterprise AI sector being overdone, which was also supported by an acceleration of growth from Snowflake when it reported Q1 26 results last week.
  • This is crucial as the market has now clearly told the software sector exactly what it needs to do to return its valuation to previous levels.
  • I have long argued that companies like Meta, Microsoft and so on, overhired before and during the pandemic, and I think they are using AI as a convenient way to get back in shape.
  • Hence, I continue not to see a job meltdown as the doomsayers predict, although the employment landscape is going to change considerably.
  • Second, RTX Spark: which is a very powerful laptop chip aimed at the developer and gaming segments of the PC market rather than taking the whole market head-on.
  • This chip consists of a Blackwell RTX GPU with 6144 CUDA cores capable of 1,000 TFLOPS with a 20-core Grace CPU (N1X) that was custom-built with MediaTek.
  • This rounds out the DGX Spark desktop product and the DGX workstation product that Nvidia launched some time ago by adding a laptop to the portfolio.
  • These products will now also support Microsoft Windows, meaning that anyone who wants to run a 1tn parameter model locally or run lots of agents can now do so with Windows rather than multiple M5 Mac Minis.
  • This product is going to be expensive, meaning that in the consumer market, where battery life and price are important, this product is unlikely to be competitive.
  • Third, AI economics: which are becoming crucially important as Jensen predicts that the cost of 1GW of AI datacentre capacity is going to climb to $80bn – $100bn, up from around $60bn where it is today.
  • Here, we saw Jensen’s usual pitch of “the more you buy, the more you save”, and given how far Vera Rubin is ahead of its competitors, he has a point.
  • Nvidia claims that it can stand up an AI data centre much faster than anyone else using its reference design, offer 2-2.5x (my estimate) more tokens per watt with better reliability and longevity.
  • Altogether, this means that revenue per gigawatt should expand substantially, and it has to, otherwise everyone is going to go bankrupt.
  • The industry standard today is $10bn/GW, but to make a decent return, I think that around $20bn/GW is needed with the latest hardware, and if Jensen is right about where costs are going, it will need to go far higher still.
  • Revenue per GW remains the key metric to watch to see who is going to make it and who will fall by the wayside.
  • This is becoming crucially important, which is why Jensen spent so much time on it.
  • Fourth, Position preservation: where Nvidia launched new models, services and reference designs all aimed at keeping partners and customers inside the Nvidia ecosystem.
  • Nemotron Ultra, Cosmos 3, Alpamayo 3, a humanoid robot reference design, where some of the launches that aim to make it easy to build AI datacentres, models, services, vehicles and robots all using Nvidia’s silicon.
  • The more companies and users base their products on Nvidia software, the more they will be wedded to Nvidia silicon, which is the aim of all of these products.
  • However, for as long as Nvidia can be at least 1 generation ahead of all of its rivals, its claim to be the best value despite being the most expensive will continue to hold water.
  • This means that the case to use Nvidia remains strong, and it is in a good position to hold onto its data centre market share as well as take share in other areas such as automotive, high-end PCs, cellular infrastructure, industrial automation, and so on.
• The net result is that this was a better-than-average keynote from Jensen with plenty of new announcements, as well as keeping his rivals in their place.
• Top of this list is Intel, whose x86 architecture was effectively called out as obsolete in both the data centre and the PC, both of which are Intel’s core markets.
• Given Intel’s recent rise, I think that the market thinks that its problems are over, but I think that short-term demand for data centre CPUs is masking reality.
• This reality will hit home when AMD announces an Arm-based CPU for both PCs and the data centre, signalling a move away from x86.
• The biggest winner is MediaTek, as its chip design expertise wins another socket with Nvidia, and Nvidia has left the consumer market open for MediaTek to attack on its own.
• This is neutral for Qualcomm as the RTX Spark is not really going up against its Snapdragon PC processors, meaning that it remains the only Arm-based supplier in the space for now.
• It also clears the way for Nvidia to continue doing what it has been doing for some time, which is growing as fast as it can build silicon, meaning a small beat and raise every time it reports results.
• This is priced into the stock now, and performance leadership has moved further down the pecking order with memory and the edge AI players (Qualcomm and MediaTek) greatly outperforming.
• I think this is likely to continue for a while yet, and so I have no intention of selling either Samsung or Qualcomm.