D2D Cont'd - Everything is an AI Trade
The sprawling impact of AI capex, a RISC V update and a look at Semis IP
Image by ChatGPT
Everything is an AI Trade Now
Since we joined Seaport, we have been publishing a fair amount and thus participating in the morning (5am!) call fairly regularly. We like to listen to the other analysts, especially those covering sectors we know nothing about. But recently, we noticed something odd. At least half the other research touches on AI. And we are not talking about other tech analysts – this is industrials, manufacturing, energy and chemicals. From what we can tell, almost every sector – other than consumer – counts AI as a major growth factor.
We are not going to speak for our colleagues, there is a lot more to their sectors than this angle, but we do think it is striking, and somewhat concerning, to see how broad the impact of AI spending is spreading into growth for so many other sectors. We are all for growth, but it is important to keep in mind that all of this is premised on capex spending from a half dozen companies, and none of them are terribly clear on their business models to justify all that spending.
We pointed this out to a long-time veteran of the industry, and he immediately made the comparison to the 1990’s dot.com bubble. Part of the reason that crash was so painful was hazy dreams of Internet spending had spread to many other sectors. We think there is something to this analogy, but conditions are different today.
For starters, in the 1990’s every company rushed to have an Internet strategy, investing heavily in projects that ultimately did not pan out, resulting in broad write-downs and layoffs. Today, the spending is much more concentrated – largely just the hyperscalers and a few venture-backed foundation model builders. The broader enterprises seem much more disciplined in their spending, lots of pilots, but no major capex unless they have a clear business case and ROI. Another major difference is that multiples are more compressed than in the 1990’s. The companies benefiting most from AI today are well established and profitable, a big change from the high fliers of the 1990’s with negative gross margins.
That being said, we do think it is worth keeping an eye on the extent to which the US economy is levered to AI spending.
RISC V’s Awkward Growth Phase
What is going on with RISC V?
There has been a lot of news in the Instruction Set Architecture (ISA) space in recent weeks, especially as it concerns RISC V – the ‘open standard’ ISA. From the outside it appears that RISC V has lost the momentum of its early years, a wave of consolidation is now sweeping the space as growth slows and the customer base shifts. And while for many, RISC V is a niche corner of the industry, its progress (or lack thereof) speaks to important changes in the broader semiconductor Intellectual Property (IP) sector.
Three items hit the news in the past two weeks:
· Global Foundries acquired IP provider MIPS.
· UK chip design service firm Codasip put itself up for sale
· Advanced packaging start-up Blue Cheetah (which is very RISC V adjacent) was acquired by Tenstorrent
Taken individually, each of these deals has commercial logic behind it, but taken together they point to a growing level of maturity for the RISC V space. Put simply, RISC V seems to becoming more important to large companies, and losing momentum among the small companies where it first gained traction. RISC V is maturing and so it makes sense that the competitive landscape is going to change. Companies like Codasip and MIPS help companies who lack complete design teams to design chips. But the large chip designers – like Qualcomm – do not need that help, and so the demand for design services is shrinking.
This leaves the open question as to the leading RISC V design service company – SiFive. They have been noticeably quiet in recent months.
Another important factor reshaping the market is the rejuvenation underway at Arm. RISC V was once seen as a potential competitive threat to Arm, but then Arm got a new management team and went public. They made some critical changes to their pricing model which made it much more affordable for small companies to sign up. This stemmed the number of start-ups flocking to RISC V because of its ostensibly ‘free’ price.
One factor that continues to work in RISC V’s favor is geopolitics. China remains a center of gravity for RISC V development, as many companies there fear losing access to Arm IP if the trade war continues. Similarly, the EU saw RISC V as a way to support their own chip design ambitions (and thus provided significant support to Codasip). This is likely to remain a factor going forward, but will need to be balanced against the substantial costs associated with using an ‘open’ ISA versus buying one off the shelf from Arm.
All of which sounds fairly gloomy for RISC V, but we think that is not quite right either. While we think the dream of an all-RISC V chip remains a dream, we do not think RISC V is going to die. Instead, we think much of the use of RISC V will move under the hood of other chips. We already see this in most RISC V deployments – those cores are now appearing in all sorts of chips, but they sit alongside Arm cores. The large chip companies are incorporating RISC V as just one more design tool. Global Foundries clearly sees this as a still viable opportunity, and we suspect that impulse was validated by customer demand. So RISC V has a future, but increasingly as IP inside more complex chips.
Semiconductor IP and Dancing Penguins
One interesting side note to our recent conversations about RISC V. Last week, we looked up several contacts who had worked with various companies in that space in the past – but all three of them had left their previous roles at three different RISC V companies. And all three now worked at Synopsys.
Synopsys is a member of the triopoly that dominates the Electronics Design Automation (EDA) space – software used to design chips. Synopsys also does a healthy business licensing other forms of IP for semis. So the migration of RISC V professionals to Synopsys is another confirmatory data point for our thesis above about the changes to RISC V.
However, we think there are other important dynamics at play here.
AI + EDA
With all the attention that AI is getting in the realm of software, many people have been bringing LLMs and transformers to the world of EDA software. Ask any of the majors and they will say they already have AI in their workflows. This comes in the form of coding assistants that can make suggestions and perform repetitive tasks. Humans still have to do most of the work, but the AI can help make it a little easier.
But AI has gotten very good at generative image creation. Surely an LLM that can draw a picture of a penguin ice skating in Rockefeller Center can draw some circuits. Not AI helping humans, but AI doing the bulk of the design itself. This possibility seems to be tantalizingly close, but not quite here yet. Three months ago, an academic team used an LLM to design an analog circuit. This was notable for a few reasons. Analog circuit design is often more art than science, with major breakthroughs achieved by humans in ‘shower moments’ or literally watering the lawn on a hot summer day. This analog circuit also looked different – clearly the work of a non-human intelligence, all organic shapes with no right angles. It now seems likely that this design still had a fair amount of human involvement, the AI did the final design but it needed a lot of hand-holding to get there. Nonetheless, teams are pushing ahead on this front across the industry, with advances announced regularly. EDA offers one area where AI seems to hold significant potential to radically alter the way work is done in a specific industry.
That being said, we do not think the EDA industry is at risk of disruption. The big three still have formidable competitive barriers in their integration with the TSMC’s PDK and process development. But there is already a growing number of new EDA start-ups emerging to tackle this new frontier.
EDA + IP
Beyond that, there is a deeper change in the semis IP space at play. Everyone in the sector is in the process of redrawing the competitive boundaries. Take the Global Foundries MIPS deal, this seems to move Global Foundries a bit further up the stack in providing design services and IP for customers. Neither of these are totally new to GF, but it still marks an incursion across the border. We can see this at Arm as well. Their Compute Sub-Systems (CSS) offering is a similar boundary crossing, not quite putting them into competition with their customers, but close to it.
The driving force behind this shift is the growing number of non-chip companies designing chips. We are getting close to the point where any company that has a clear software stack can design the compute logic of a chip with a fairly small team. There is still a big jump between that and production silicon, but that gap is the opportunity for the IP companies to fill. It is not clear how many companies still want to design custom silicon, but it is getting much easier to do so. Ultimately, we think the real test will come with the automotive industry. Talk to any of these IP providers and it is clear they see opportunity in two areas – AI and automotive. AI is already hotly contested, but it is not clear how far the auto companies will go in designing their own silicon. PRC auto OEMs are pushing down this path hard. EU auto OEMs seem to be licensing chips and software from PRC companies. And US OEMs seem to be deer in the headlights, frozen in indecision. How that landscape shapes up will be a major deciding function in the ultimate fate of the semis IP landscape.