AI Chip Supply Chain Boom: Nvidia Orders, Terafab, and Photonic Memory

LLM inference at scale runs into a hard physical constraint: the key-value cache that keeps a running conversation's context alive can demand tens of terabytes of memory at hundreds of gigabits per second, and no existing memory tier has delivered both capacity and bandwidth at once. Traditional electrical signaling has confined high-bandwidth memory to within about 1.5 millimeters of the GPU package. Marvell's new Photonic Fabric Memory Module replaces that copper tether with optical fiber, letting HBM and DDR memory sit up to 50 meters away while still contributing to a shared, poolable memory tier spanning multiple XPUs and racks ^[1]. The module can offload up to 32TB of warm KV cache and, per Marvell, lifts token throughput by 2-3x within existing data center power and space budgets ^[1]. The shared-memory framing matters as much as the distance: instead of memory being locked to a single accelerator, racks can now draw from a common pool ^[3]. Engineers caution the practical sweet spot is closer to 25-30 meters rather than the full 50-meter spec, since latency creeps up with fiber distance even as raw access latency stays in the hundreds of nanoseconds ^[2]. Marvell's Will Chu frames this as infrastructure 'moving beyond isolated servers to systems where compute, memory and connectivity operate seamlessly together' ^[1], while independent analyst Alan Weckel argues memory capacity, bandwidth, latency and data movement - not raw compute - are becoming the primary constraint on AI performance ^[1]. That reframing is the real news: Marvell isn't just selling a faster part, it's proposing that memory placement and compute placement no longer have to be the same decision.

Elon Musk says Terafab - the Tesla/SpaceX chip-manufacturing venture - will produce output from a single building valuable enough to 'rank in the top 20 of country GDPs' ^[4][5]. That's a striking claim: the lower boundary for a top-20 economy is roughly $1.15 trillion, while the entire global semiconductor market is forecast at only about $1.5 trillion in 2026 ^[6]. Musk hasn't shared the math behind the comparison, and the reporting around Terafab itself is a patchwork of conflicting numbers. Location: some outlets place it in Grimes County, Texas, next to Gigafactory Texas ^[4], while others describe the Seaholm Power Plant site in Austin ^[6][7]. Cost: figures range from a $16.8 billion 'first phase' ^[4]to a $20-25 billion initial estimate, a $55 billion investment filing, and an up to $119 billion total buildout ^[6]. What's consistent across sources is the physical scale - over 100 million square feet - and an initial target of 100,000 wafers per month scaling toward 1 million wafers per month within a decade ^[4][6]. Intel joined as a foundry/technology partner in April 2026 to help implement an advanced 14A process ^[8], even as Terafab's backers insist Tesla and SpaceX, not Intel or TSMC, will run the facility ^[6]. Musk's own justification is blunt: 'we either build the Terafab or we don't have the chips,' since he says existing global fab output covers only about 2% of what his companies need ^[7]. Until the cost, location, and output figures converge across reporting, the GDP comparison reads more like a recruiting pitch - aimed partly at the roughly ten South Korean chip suppliers Tesla and SpaceX are courting, more than half of whom have already signed NDAs ^[4]- than a verified forecast.

ByteDance, Alibaba, and Tencent have ordered more than 3 million of Nvidia's RTX PRO 5500 cards, worth over $35 billion at roughly $11,400 each, with total Chinese orders exceeding 6 million units ^[9]. The RTX PRO 5500 is notable less for its specs than for its legal status: it's a workstation chip rather than a data-center accelerator, which makes it among the most powerful Nvidia silicon still permitted for Chinese buyers under US export curbs that block the more advanced B300 and H200 models ^[10]. Nvidia has told partners they may receive only 5-10% of the volumes requested, capping global production at roughly 3 million units in 2027, with early samples due in December 2026 and commercial deliveries in Q1 2027 ^[10]. Starting in 2027, Nvidia is also expected to redirect much of its TSMC capacity toward the next-generation Rubin platform rather than chips like the RTX PRO 5500 ^[10]- meaning this China order may be as much a one-time capacity grab as a durable revenue stream. Tellingly, Nvidia's stock barely moved on the news, trading only marginally higher even as it remained down more than 9% from recent weeks ^[11]; Nvidia's own CFO has said less than 1% of data-center revenue comes from China and that Chinese sales were never baked into guidance, against a projected $108 billion in Q3 FY2027 revenue ^[11]. Marvell's photonic memory story has drawn a louder market reaction by comparison: its stock reportedly jumped as much as 29.1% after the AI Memory Infrastructure Portfolio launch ^[12]and rallied another 6% weeks later on its 2nm optical interconnect showcase ^[13], with analysts framing the move as Marvell 'architecting' the solution to the industry's optical memory wall ^[14]. The contrast is instructive: Nvidia's order is large but politically contingent and margin-thin; Marvell's smaller headline numbers are moving a stock because they represent a structural shift in how AI infrastructure gets built, not just who gets to buy it.

Dave Lazovsky, the Marvell executive associated with the Photonic Fabric line, isn't a lifelong Marvell employee - he's Executive Vice President and General Manager of Marvell's Data Center Networking Business Group, a role he stepped into after Marvell acquired Celestial AI, where he was co-founder and CEO ^[15]. At Celestial AI, Lazovsky helped pioneer the optical scale-up connectivity that now underpins Marvell's Photonic Fabric line, bringing roughly three decades of semiconductor-industry experience to the role ^[15]. That provenance matters for reading the competitive map: Marvell didn't build photonic memory disaggregation purely in-house on a product roadmap - it acquired a startup's core IP and is now shipping it as a flagship data-center platform, the same acquire-and-scale pattern increasingly common across AI infrastructure. It also means the technology's credibility rests partly on Celestial AI's prior optical-interconnect work rather than purely on Marvell's own R&D history, a distinction worth keeping in mind given how aggressively the photonic-memory narrative is now being marketed to investors.