AMD's 245x Multiple, Firmus IPO Collapse & GlobalFoundries' $2B Packaging Deal

Jamie Cole (0:00)

AI Hardware & Chips, Daily News.

I'm Jamie Cole, thanks for joining me.

Today, AMD's Trillion Dollar Balancing Act.

AMD's data center revenue just doubled in a single quarter, and yet the company's valuation is the thing that's keeping analysts up at night. AMD is trading at 245 times trailing earnings. That's not a typo.

After a nearly 200% gain year to date, the bull thesis now requires almost flawless execution. $20 in earnings per share, a successful Helios GTU ramp, and zero slippage on commitments from OpenAI, Meta, and Anthropic. BNP Paribas has a $960 price target, but even that requires AMD to hit 48 times multiple on earnings it hasn't posted yet. The key implication is simple. The stock isn't pricing in what AMD has done. It's pricing in everything AMD needs to do next.

While AMD's growth narrative holds together for now, the public markets are sending a different signal elsewhere. Firmus, an Nvidia-backed Australian AI infrastructure firm, just pulled a $5 billion IPO. The company had been valued at just over $30 billion, and had raised $2 billion in a private round led by Nvidia back in August. The withdrawal was attributed to market volatility. That matters because Firmus isn't a speculative startup.

It has real backing and real revenue expectations. Its decision to go back to private capital, rather than face public market scrutiny, suggests that appetite for AI hardware valuations at scale is thinner than the private funding rounds implied. Whether this is specific to Firmus, or a broader chill on AI infrastructure IPOs is still unresolved. Either way, it's the first significant pullback in this cycle, and the timing is hard to ignore. Away from public market noise, something more structurally important just became contractual. GlobalFoundries has signed a $2 billion multi-year deal to produce silicon interposes at its Malta, New York facility. Volume production is targeting the first half of 2028, supporting advanced packaging for AI and high-performance compute customers. Advanced packaging has been an architectural talking point for years. This deal makes it a binding supply agreement. The signal here is that chiplet-based AI accelerators are moving beyond roadmap slides and into capacity commitments. GlobalFoundries is locking in US-based packaging capacity through the end of this decade, which aligns directly with the broader push for geographically diversified AI infrastructure.

There's a risk on the other side, if AI chiplet demand moderates after 2028, that capacity becomes a stranded investment. But right now, the directional bet looks well-timed.

Here's the thing, HPE has launched four new ProLiant Gen13 servers, all built on AMD's 6th generation EPYC processors. The DL525 offers 256 cores in a single socket.

The DL585A supports 8 GPUs and targets inference workloads specifically. This is a significant OEM vote of confidence, and it removes a layer of execution uncertainty that was previously sitting on AMD's roadmap. Intel's data center revenue decline makes the contrast harder to miss. AMD grew data center revenue by 107% in the most recent quarter. HPE's full pivot to EPYC reflects that shift in commercial reality. AMD is also quietly fragmenting its service CPU line. The upcoming Verano processor, based on Zen 6 architecture, uses a new SB1 socket, carries 72 cores and runs on a 24-channel Lpddr5X memory subsystem. That's a deliberately different design from Venice, AMD's general purpose server line. The implication is that AMD is building separate CPU architectures, optimized for AI inference, rather than trying to unify everything under one socket. That's a sign of market maturity. It's also a source of ecosystem complexity. How broadly OEMs will adopt the AI specific SB1 socket alongside Venice remains an open question, heading into Verano's planned second half 2027 launch. Consider this. One more development worth tracking. Qualcomm CEO is pushing back on the idea that AI makes smartphones obsolete. The argument is that AI agents running personal context require on-device inference, not cloud round trips. To back that thesis, Qualcomm acquired Modular, positioning the move as an alternative software stack for developers who don't want to build inside Nvidia's ecosystem. The important distinction is that this isn't just a hardware play. It's a developer platform bet. If on-device AI gains traction, Qualcomm wins at the edge. If cloud AI remains dominant, the addressable market for that bet stays narrow. The through line across all of this is AMD's balancing act. Strong fundamentals, real customer commitments, and OEM confidence, but a multiple that assumes everything works on schedule, at scale, with no export control surprises. Watch Helios' production milestones. Watch whether Firmus remains an isolated case, or the start of a pattern. And watch how many OEMs follow HPE into the AMD ecosystem, before Intel responds. Thanks for listening. This podcast was built using AI technology, a YesWe production.