Arm – Feeling Chipper

Arm twists the knife on Intel.

• Arm has launched its first silicon chip, which already has widespread support and is clearly targeted at Intel and AMD and capturing a much larger part of the x86 market than it has been able to historically.  
• At its Arm Everywhere event, Arm launched its first silicon chip, the Arm AGI CPU, which is targeted at the growing CPU opportunity in the data centre.
• This completes the Arm line-up for the data centre, which begins with IP blocks, is extended with compute subsystems (CSS), which puts those IP blocks together, and AGI CPU that is now available as a full chip for those that don’t want to design the system themselves.
• Arm has launched this product as it already has robust demand from clients (beginning with Meta) and because it has the potential to greatly increase the opportunity for Arm to grow revenue.
• This is because the move to using agents to execute tasks and carry out research, rather than simply to generate data, places much more compute demand on the CPU rather than the GPU.
• It is on this basis that Arm thinks that 30m CPU cores per GW of AI compute will go to 120m to ensure that the CPU is not a bottleneck holding up the data centre.
• All things being equal, this would mean a 4x increase in power requirement, which, given the increasing constraints on energy, rapidly becomes a problem.
• This has always been Arm’s key selling point, and this time around, Arm takes a leaf out of Qualcomm’s book and has designed its product to offer the same performance as x86 but consuming half the power.
• In the data centre, this is more relevantly expressed as twice the performance for the same amount of power.
• This was emphasised time and again right the way through the session, and it was against x86 that all of the comparisons were made.
• Hence, it is against x86 that Arm is competing (meaning Intel and AMD) and not so much against its customers, such as Qualcomm, MediaTek and Broadcom.
• The product has two main selling points:
  • First, performance per watt: where Arm claims a 2x performance at the same power, which is ideal for a data centre which is going to need to increase CPUs per GW by 4x.
  • This advantage has long been known in the data centre, but the problem of legacy software has made it very difficult to use Arm, as all the old software had to be ported to the new instruction set.
  • However, now that power is so much more of a problem than it was, and the data centre is going through a transition, this means that a lot of the software has to be rewritten anyway.
  • In the data centre, power efficiency has double the impact as not only does it use less power (meaning a lower bill), but also less cooling is needed, reducing costs further.
  • Hence, this is the optimal time to switch to a new instruction set, and this is why I suspect Arm is receiving such broad support across the industry.
  • Second, industry support: which is going to be critical to the success of this new venture from Arm.
  • The chip was designed in partnership with Meta, but OpenAI has also said that it will use it within its infrastructure.
  • Both companies appeared on stage to extol the virtues of the chip and said that they will use but refrained from saying how much they are going to buy.
  • Aside from committed customers, Arm is also able to point to 60 partners who will be supporting the new chip, several of whom also recorded a video endorsement.
  • This is what impressed me most about this launch as with most of the industry behind it, Arm has a good chance of winning traction and making a success of this new business line.
• However, not everyone is happy, as some of Arm’s biggest customers were nowhere to be seen at this event, and this will crystallise fears that Arm intends to compete with its customers.
• The counter to this is that Arm is only targeting x86 with this product line, and that the market is becoming so large that there is plenty of space for everyone, but not everyone will agree.
• For example, Qualcomm is currently focused on NPU-based accelerators while MediaTek is working with Nvidia and others using its expertise in certain areas to help them to create custom CPUs.
• On balance, when I look at the positioning of this product, I do not see it as interfering with any of the existing business lines of its clients, but one could argue that customers are planning to address this opportunity where Arm will now be a competitor.
• This means that even though there is no direct competition, Arm will have to tread carefully to ensure that the ecosystem is happy and that the incentive to move to RISC-V is minimised.
• The net result is that moving into selling chips rather than IP greatly increases Arm’s data centre addressable market, which it estimates has gone from $3bn (IP and CSS) to $100bn by 2030 (IP, CSS and chips).
• This is why this session was pretty well received by the market which sent the shares up by 7% in after-hours trading.
• However, it is more terrible news for Intel, which is under assault from every direction and seems to have no idea what to do about it.
• This is just another sign that x86 is obsolete, and I am looking for AMD to launch data centre and PC chips based on Arm for confirmation of this view.
• There remains no price at which I would want to buy Intel.