Americans invented Beatlemania in 1964, which was bittersweet for the British, who had invented it in 1962. Something similar is happening now to Arm Holdings.
The Cambridge-headquartered chip designer has doubled in value after posting quarterly results last week. Since its Nasdaq float in September, the stock is up 192 per cent. Its market cap at Monday’s close of $152.8bn would put it third by weight in the FTSE 100, ahead of HSBC and just behind Shell, if it were London listed, which it once was but now is not.
The straightforward explanation for the move is that Arm’s auditioning to join the Magnificent Seven. AI gets 19 mentions in Arm’s third-quarter investor presentation and was the leitmotif of its conference call. Everyone who’s anyone in AI is an Arm licensee, including Nvidia, which in 2020 agreed to buy the company for $40bn then was forced by the FTC to abandon the plan.
Arm’s an option if you’re looking for a name that’s at least tangential to all the themes A16z types blog about, like AI-optimised datacentres, AI acceleration, AI open standards, AI mixed computing, edge AI and custom AI inference. At least some of the $7tn that AI’s carnival barker Sam Altman says he’s raising would end up at 110 Fulbourn Road, Cambridge, and even a small percentage of $7tn is still a big number.
To those of us who were writing about Arm before the dotcom bubble, there’s a lot here that’s familiar. There’s also a lot around that can encourage hope built on misunderstanding, such as this quote from analyst Dylan Patel of SemiAnalysis:
Patel says [Arm] has a huge amount of room to increase the royalty rates it charges device makers that use its chip designs. He notes Arm charges phonemakers something on the order of 50 cents a phone for its mobile CPU designs. Qualcomm’s royalties for using its radio chips in a high-end phone might run something like $13.
Every management meeting between Arm’s London float in 1998 and its purchase by SoftBank in 2016 had a section on royalty rates. The gist was that even though Arm had a 100 per cent monopoly in areas like smartphone CPUs, ubiquity did not mean pricing power.
Arm has remained the chip industry’s de facto standard by being neutral. It offers universal, standardised ways for companies to cut development costs while refusing to discriminate by price. The rule of thumb has always been that royalties are 5 per cent of the chip price. And while Arm’s per-unit royalties have nearly always trended gradually higher, it’s only because licensees have built hardware using its more powerful architectures.
That’s not to suggest that the 5-per-cent rule can’t change. A recent initiative is for Arm to sell subsystem licences (chip designs with paths for GPU and memory already integrated, basically) that help automakers and appliance companies connect everything up. But Arm’s expectation is that subsystem royalty rates can rise to 7 or 8 per cent of chip cost, so the improvement per unit is still measured in cents not dollars.
Arm posted record royalties in the third quarter, with revenue up 12 per cent quarter on quarter to $470mn, as customers moved from its V8 architecture to V9, which was launched a year ago.
A trend in Arm’s past life was that, because hardware upgrade schedules and demand for consumer electronics are both highly cyclical, its licence sales would disappoint whenever royalties impressed. And because for most companies licensing is a big one-off cost, their lumpiness was a recurring blight on Arm’s numbers.
Arm’s spread into more diverse markets could help level out this cyclicality. Another change since the SoftBank buyout is around how Arm charges licensees. Previously, everyone paid upfront for specific licences. Now the company runs an annual subscription model that gives big customers access to everything, with discounts available for smaller operators. There’s also a licence premium levied on the likes of Apple and Qualcomm, who customise off-the-shelf Arm architectures and are likely to pay lower per-unit royalties as a result.
Is it working? Hard to say. Third-quarter licensing revenues (43 per cent of sales) were down 8 per cent quarter-on-quarter at $354mn. It’s not that much for a company whose designs are written into the product roadmaps of Alphabet, Samsung, Apple, Microsoft, TSMC, Intel, Nvidia, TSMC and Amazon (all of whom volunteered to anchor the Nasdaq IPO, another symptom of Arm’s closely policed industry neutrality).
The main contributor to Arm’s better-than-expected third quarter was China, which accounts for a quarter of group sales and where revenue grew by 28 per cent from Q2.
The company doesn’t give detail for individual territories but it’s likely that Chinese royalties were boosted by fancier handsets such as the Vivo X100, one of the first to include generative AI functions. Strip out China and group revenue would probably have been down sequentially.
After the surprise revival of demand for premium smartphones in China, investors are betting on the same trend worldwide. That’d be great for Arm, though is also a reminder of unresolved problems with its Chinese joint venture. Royalties from the semi-autonomous JV still need to be booked as related-party revenue.
Eventually, every article on Arm has to mention valuation. The stock’s forward PE of 124, falling to 78 by 2026, is clearly a bit nuts. It’s twice as rich as sort-of peers like Cadence Design Systems and Synopsis. Back in 2015, when optimism was running high around how the V8 architecture would boost royalties, the stock was on an ex-cash forward PE of 32.
But not everything can be tied back to Arm’s former life as the London market’s tech flag-bearer. For one thing, management seems to have traded English understatement for American optimism.
The guidance at IPO was for sales growth to accelerate from 22 per cent in 2025 and 25 per cent in 2026, from 11 per cent growth this year. (The consensus since risen to imply 2024e growth of 18 per cent.) Given the maturity of the smartphone market and Arm’s long history of tepid margin expansion, that’s going to require a hell of a lot of data centres, etc. Yet the current share price would imply that targets have been lowballed.
Is generative AI going to be transformational for Arm as investors expect? Maybe, though the ways it’ll happen are not easy to predict. As highlighted often, Arm’s niche of power-efficient chip architectures are too slow for gen-AI training (the industry term for building the database) but might be great for inference (pulling data from the database). Then again, so might lots of other flavours of chip. As Morgan Stanley writes:
The main desktop CPU architectures — x86, Arm, and maybe RISC-V too — have decades of software toolchain investment behind them, have huge developer communities and mature development tools. These all eclipse the scale of NVIDIA’s community. But what’s sometimes overlooked is that these CPUs can be much cheaper per logic operation than a GPU. As such, when we see the pivot to inference — which will trade-off low latency over throughput, certainly in edge-AI — we may see a ramp in the $-sales in Arm’s Infrastructure business.
Risc-V, mentioned above, is an alternative architecture for low-power chips that’s open-source and free to use. Though dwarfed by the Arm ecosystem, Risc-V designs can be found in products including Google’s Pixel handsets and Seagate data servers. Other developers take their own route, such as Tesla, which says its Dojo AI chip project is entirely home-brewed.
That makes the Arm investment case not dissimilar to the one we were obsessing over a decade ago. Chips ship in huge volumes but the main competition comes from companies’ own R&D labs, so pricing power is weak. What matters for Arm forecasts a few years out is that most abused of metrics, Total Available Market. Pick a number, in other words.
In the meantime, it’s worth remembering that the post IPO lock-up on insiders selling will expire on March 12, and 90 per cent of Arm shares are still owned by SoftBank. Arm-mania is raging for the moment, but there’s an obvious candidate for who will be the band’s Yoko.
Read the full article here