The Circuit: Memory's New Floor, Balance-Sheet-as-a-Service, and Why Nobody Can Agree on What a Token Is Worth
Podcast roundup following Micron earnings, HPE's networking investor day, and OpenAI's October 6, 2025 DevDay
The latest episode of The Circuit, hosted by Ben Bajarin and Jay Goldberg, used a week stacked with catalysts, Micron's earnings, an HPE networking investor event, and OpenAI's DevDay, to make a broader argument that the market still hasn't settled on how to model the structural shift underway in memory, server financing, and AI token economics. The conversation ranged across Micron, HPE, NVIDIA, Broadcom and AMD, but the throughline was that old frameworks for valuing these businesses are breaking down faster than Wall Street's models are being rewritten.
Memory's Floor Has Moved, But Has the Cycle?
Micron's quarter was clean, numbers, guide, and tone all landed well, yet the stock finished the week roughly flat, which Bajarin and Goldberg read as evidence that the market remains deeply split on what memory is actually worth now. Bajarin laid out the core tension: the addressable memory market has clearly re-rated, from a historical peak near $100 billion to a modeled $600 billion to $800 billion depending on where the new ASP floor settles, yet investors are still pricing in a hard crash arriving around 2029. "That is what's being modeled," Goldberg said, "demand destruction and a high degree of margin compression." Bajarin pushed back on the assumption that cyclicality has to end in a crash rather than oscillate around a permanently higher floor, calling the debate the real swing factor for Micron and its peers heading into 2028, when DRAM and NAND bit growth, running in the low-to-mid twenties percent through 2027-2028, could catch up to the furious capacity expansion now underway, creating what Goldberg termed a potential "supply-demand equilibrium" that the market could start pricing pessimistically well before it arrives.
China's CXMT Is the Wildcard Nobody Is Modeling Correctly
Both hosts flagged Chinese memory maker CXMT's accelerated capacity targets as an underappreciated risk to the 2028 supply picture. The real edge Chinese memory makers hold isn't unconstrained access to capital or technology, it's construction speed. "It takes two and a half, three years to build a memory fab in the US, Korea, or Japan," Goldberg noted, "it takes 18 months in China." That gap means a meaningful chunk of new global bit supply could land earlier and faster than Western capacity plans assume, compressing the supply-demand timeline that bulls are counting on to delay any pricing reset.
The Case for "Trailing-Edge" Memory as a TSMC-Style Legacy Node
The most structurally interesting idea in the episode was Bajarin's argument that memory may be evolving toward a foundry-like model where older nodes, DDR4 in particular, stay monetized indefinitely rather than being retired and replaced, as has historically been the case. With bit growth for AI accelerators effectively unbounded ("if you talk to anybody who makes logic around AI accelerators, they just want more, it's memory hungry," Bajarin said), he argued there's a rational case for memory makers to keep legacy clean-room capacity running to serve phones, PCs and CXL-based disaggregated compute, while directing new capacity and R&D toward HBM. Goldberg was far less convinced the memory makers have actually internalized this logic, pointing out that the economics currently favor cannibalizing trailing-edge capacity for HBM: a wafer that produces four units of commodity memory might instead yield one HBM die sold at roughly 100 times the price. "The math works very heavily in the favor of going, just throwing everything they can at HBM," he said, adding that Tesla's recent decision to de-spec memory in its robots due to supply constraints is an early warning sign of what happens to non-data-center customers when capacity gets starved. Goldberg also raised a separate demand-side risk largely absent from current memory models: smartphone upgrade cycles may be structurally lengthening as devices get pricier, with one forecaster telling him the industry could face "a decade of no growth in mobile phones" absent an AI-phone catalyst.
HPE Wants to Be Known as a Networking Company, But What's the IP?
Bajarin's recap of HPE's networking investor event surfaced a sharper, more skeptical read on the Juniper-driven pivot than the headline optics suggest. HPE executives are explicitly positioning the company as a networking business first, with infrastructure and software pulled through behind it, driven partly by enterprise customers' fear of agentic-AI-era security vulnerabilities in aging campus networking gear. The company showcased its Helios rack, built around Broadcom Tomahawk switches running UALink-over-Ethernet for scale-up domain switching, which Bajarin described as containing "the most insane amount of copper ethernet cables I've ever seen in one single compute tray." But the engineering showmanship raised an uncomfortable question that multiple sell-side analysts, including Morgan Stanley's Erik Woodring, were asking on the show floor: what exactly is HPE's proprietary IP here, given the switches are Broadcom's and the compute is AMD's or NVIDIA's. Goldberg, who covered Juniper for years, was blunt about his own skepticism of the networking technology underpinning the strategy, while Bajarin conceded the company's "Switzerland" positioning, supporting NVLink, Ethernet, and any accelerator vendor, may ultimately mean HPE's value is less proprietary IP and more world-class systems integration.
Balance-Sheet-as-a-Service Becomes a Formal Slide, Not a Footnote
One of the sharpest observations of the episode centered on NVIDIA's non-deal roadshow deck, which circulated to at least six sell-side banks this week and, for the first time, included a slide explicitly framing balance-sheet usage as a growth strategy rather than a risk disclosure. Goldberg noted the slide was essentially a repackaged image of NVIDIA's 10-K obligations footnotes, the same image bears have used to highlight ballooning commitments. "It's like that meme from The Big Short," he said, "why are they confessing? They're not confessing, they're bragging." Layered on top of this was news that Broadcom has committed to lend Anthropic up to $42 billion, per a Reuters report on a preliminary Anthropic IPO filing, tied to Google TPU production. Bajarin pressed the more interesting strategic question: why is it the fabless chip designers, NVIDIA and Broadcom, rather than foundries like TSMC, stepping up to underwrite customer capacity. Goldberg's answer was that TSMC sits too far up the supply chain to read true end demand ("it's in the news when someone from Google or Sam Altman goes to Taiwan to talk to TSMC, that's how rare it is") and that NVIDIA and Broadcom, both of which carry materially higher gross margins than TSMC, are simply better positioned to monetize and price that AI demand directly. The discussion also drew a pointed comparison to Hock Tan's history of running Broadcom "like a private equity fund," with Bajarin arguing Broadcom's entry into vendor financing is itself a signal that sophisticated capital allocators see enough structural durability in AI infrastructure spending to take on the risk. Goldberg was more cautious, framing the dynamic as NVIDIA forcing every competitor, AMD included, to match its balance sheet, networking, and systems capabilities just to stay competitive, a race AMD's roughly $10 billion cash position cannot realistically win against NVIDIA's $60 billion-plus cash balance and vastly larger free cash flow generation. "If you told me five years ago the central competitive dynamic in semiconductors was use of balance sheet and taking on debt," Goldberg said, "I wouldn't have believed you."
Tokenomics: Pricing Finally Becomes Quantifiable
Goldberg used the episode to introduce what he sees as a maturing analytical framework he calls tokenomics, arguing that enough pricing data now exists across input versus output tokens, reasoning model premiums, and inference speed tiers to move AI unit-economics analysis from hand-wavy commentary into hard-dollar modeling. He noted that the reasoning-model pricing premium labs charged roughly a year ago has effectively disappeared, now bundled into frontier model pricing as a standard feature rather than a monetizable differentiator, while input and output token pricing continue to diverge because the underlying compute costs are fundamentally different. This framework, he argued, also explains strategic moves like NVIDIA's reported $20 billion acquisition of Groq. The conversation connected directly to OpenAI's DevDay announcement of an "Ultrafast" tier running at roughly 250 tokens per second, which Bajarin noted was initially mistaken for Cerebras-powered inference before it emerged the speed-up came from a low-batch-optimized configuration of standard NVIDIA GPUs, not a specialized accelerator. Both hosts agreed this points toward a near-term future of segmented, speed- and model-based pricing tiers rather than a simple binary of "premium" versus "commodity" tokens, with Goldberg predicting enterprise customers will increasingly need their own tokenomics models simply to negotiate effectively with labs now sophisticated enough to employ in-house economists on pricing strategy.