In partnership with
Elon Musk announced the most ambitious chip factory in history this week. There are caveats.
He has no timeline, no EUV lithography machines, and no semiconductor manufacturing experience. He did, however, say he could eat a cheeseburger and smoke a cigar in the clean room.
At 2nm, individual transistors are smaller than a virus, and one speck of dust destroys a wafer.
But sure, light up.
I'm Ben Baldieri, and every week I break down the moves shaping GPU compute, AI infrastructure, and the data centres that power it all.
Here's what's inside this week:
Let's get into it.
The GPU Audio Companion Issue #99
Want the GPU breakdown without the reading? The Audio Companion does it for you, but only if you’re subscribed. If you can’t see it below, click here to fix that.
Today's issue is brought to you by Rafay
The GPU lives on sponsorships. If you value independent, no-fluff analysis of the AI infrastructure market, the sponsors are who make it possible. I'd appreciate you checking them out.
How to Build an AI Factory with Rafay
Organisations are investing heavily in GPUs, but infrastructure alone doesn’t create an AI factory. That’s why Rafay put together this paper exploring the architectural and operational model required to transform raw GPU infrastructure into a self-service AI platform. Click the button below to learn how leading companies operationalise AI infrastructure with governance, developer self-service, and production-ready AI workflows.
Musk Announces $25 Billion Chip Factory. The Caveats Are Enormous.
Elon Musk says he's going to build the world's most advanced chip factory from scratch. The semiconductor industry would like a word.
Tesla and SpaceX have announced Terafab, a $20-25 billion semiconductor fabrication complex in Austin, Texas, jointly owned by Tesla, SpaceX, and xAI. The facility targets 2nm process technology and would consolidate chip design, lithography, fabrication, memory production, advanced packaging, and testing under one roof. Two chip types are planned: an inference processor for Tesla vehicles and Optimus robots, and a radiation-hardened chip for SpaceX satellites and orbital AI.
Musk claims global chip output meets only 2-3% of his companies' future needs and said he told TSMC, Samsung, and Micron directly: "We will buy all of their chips." No construction timeline was given. Musk described it as "the most epic chip building exercise in history by far."
Why this matters:
The caveats start with physics and get worse from there. TSMC spent decades and $165 billion building the capability to manufacture at 2nm. Its Arizona fabs alone won't reach 2nm production until 2029. The process requires a fundamental transistor architecture shift from FinFET to GAAFET, where minor deviations at any stage can collapse yields.
TSMC has roughly 50,000 engineers who do nothing but semiconductor manufacturing. Tesla has zero. ASML's EUV lithography machines, the only tools on earth capable of producing at this node, cost $350 million+ each with multi-year waiting lists. Major customers have booked production slots years in advance. Tesla's CFO acknowledged the full cost isn't yet in the capex plan. Morgan Stanley estimates $35-40 billion and no chips before 2028. Tesla made less than $4 billion in profit last year. The global pool of qualified fab construction managers numbers in the hundreds. Tesla is currently advertising to hire one.
And then there's the cheeseburger question. Musk said in January he could "eat a cheeseburger and smoke a cigar" in the fab, dismissing conventional clean room requirements. The proposed site is at Giga Texas, next to stamping machines that shake the ground. For context, Musk's 4680 battery cell programme, announced at Battery Day 2020 with a promise of 10GWh within a year, is at roughly 2% of original volume targets years later. Chip fabrication at 2nm is orders of magnitude more complex than battery cells.
Elliott Takes Multi-Billion-Dollar Stake in Synopsys
The activist investor that reshapes companies just bought into the company that designs every AI chip.
Elliott Investment Management has taken a multi-billion-dollar stake in Synopsys, the electronic design automation (EDA) company whose tools are used to design virtually every advanced semiconductor on earth. Elliott's statement: "There is a clear opportunity for Synopsys's financial performance to more fully reflect the value it delivers." Synopsys customers include AMD, Arm, Intel, Microsoft, Samsung, TSMC, and NVIDIA, the latter of which invested $2 billion in Synopsys in December 2025 as part of a multi-year chip engineering partnership.
Why this matters:
Synopsys and Cadence are the duopoly that sits underneath every chip announcement in this newsletter. Every NVIDIA GPU, every AMD Instinct, every Meta MTIA, every Arm AGI CPU (this issue) was designed using their tools. Elliott pushing for greater profitability at a company with that kind of structural positioning means higher licensing costs flowing through to every chipmaker and every customer downstream. When the tool vendor gets squeezed by an activist, the entire supply chain feels it.
NVIDIA invested $2 billion in Synopsys three months ago. Elliott arriving now creates a tension between NVIDIA's interest in keeping its design partner aligned on long-term R&D, and Elliott's interest in extracting near-term margin. Activists at EDA companies don't typically push for more R&D spending. They push for pricing discipline and cost cuts. The chipmakers who depend on Synopsys tools will be watching closely.
The broader pattern: infrastructure-layer companies that were invisible two years ago are attracting capital at scale. CoolIT (liquid cooling) sold for $4.75 billion last week. Nexthop AI (networking) raised $500 million. Now Elliott is deploying billions into chip design tools. Every layer of the AI infrastructure stack, from power to cooling to networking to silicon design, is being repriced as strategically critical.
Nscale's Predecessor Defaulted on Loans, Lost $102 Million on $19 Million in Revenue
The third piece of critical Nscale journalism in three weeks. This time it's the FT.
FT Alphaville reports that Arkon Energy, the miner from which Nscale was born in late 2024, defaulted on two loans from Sandton Capital Partners, a New York distressed credit investor. Arkon pulled in $19.2 million in revenue in 2024 against a $102 million loss. Staff costs alone were $19.7 million. Sandton lent roughly $110 million across multiple facilities at interest rates as high as 17.5%, with steep discounts: $37 million extended in exchange for $44 million owed, before interest. Arkon triggered a default in December 2023, received a waiver, then defaulted again on principal repayments that were eventually repaid from the Series B. Sandton led the Series A and participated in every subsequent round. It holds warrants equivalent to a significant equity stake and is a top-five investor.
Why this matters:
Issue #72: Sifted lawsuit and governance questions. Issue #96: $2b Series C, Guardian investigation, scaffolding yard in Essex, government "not playing an active role in auditing." Issue #97: 8GW West Virginia acquisition. Now: FT reveals the predecessor company borrowed from a distressed credit specialist at 17.5% and defaulted on repayments eighteen months before the $14.6 billion valuation. The capital trajectory is extraordinary. So is the gap between the headline numbers and the corporate history underneath them.
Sandton lent at distressed rates, received warrants, led the Series A, participated in every raise, has a board seat, and is a top-five investor. The defaulted loans were repaid from equity that Sandton itself participated in. That's a closed loop. The FT notes Sandton was "firmly in the boat, and understandably didn't want to rock it." None of this is illegal. All of it is worth understanding when evaluating a company now valued at $14.6 billion and building 8GW of US infrastructure for Microsoft.
Nscale's response: the defaults were "an accounting technicality" because rollover paperwork was delayed while everyone focused on closing the Series A. That framing may be accurate. It may also be exactly the kind of explanation the Guardian's investigation suggested governments should be auditing rather than accepting at face value.
OpenAI Shuts Down Sora, Retreats from Video Generation
Six months after a million downloads in five days, OpenAI killed Sora.
OpenAI is closing its video generation app, shelving Instant Checkout, and consolidating apps into a single desktop product. A planned $1 billion Disney investment never closed. Disney said it "respects OpenAI's decision to exit the video generation business." CEO of Applications Fidji Simo told staff the company is "orienting aggressively" toward high-productivity use cases and competing in enterprise, where Anthropic has built a significant business with Claude.
Why this matters:
Sora was always a consumer spectacle and a convenient fundraising tool, not a revenue driver.
The Disney deal's (surprise) collapse is the tell: even a billion-dollar partnership with the world's most valuable IP library couldn't make the economics work. OpenAI is retreating to enterprise and productivity, the exact territory where Anthropic has been gaining.
The consumer AI era Sora represented is being deprioritised by its own creator in favour of the workloads that generate revenue, though how enterprise customers feel about models optimised for engagement and sycophancy over business outcomes remains to be seen.
Meta Partners with Arm to Co-Develop Custom Data Centre CPUs
Meta just added a fifth silicon partner. This time it's the CPU.
Meta and Arm will co-develop multiple generations of data centre CPUs for AI workloads. The first, the Arm AGI CPU, is Arm's first data centre CPU designed for AI, optimised to work alongside Meta's MTIA accelerators. Board and rack designs will be published under the Open Compute Project.
Why this matters:
Five silicon partners: NVIDIA, AMD, Broadcom (MTIA co-design + standalone accelerators (Issue #96), MTIA in-house, and now Arm for CPUs. No other company on earth is running this many parallel hardware programmes. Meta is designing every component of its server stack with a different partner.
The CPU was the missing piece. NVIDIA ships Vera CPUs. Google has custom TPU host processors. Amazon has Graviton. Meta was running commodity CPUs until now. Purpose-built silicon from CPU to accelerator to rack means Meta can co-design the entire system as one integrated unit. That's the same vertical integration thesis driving NVIDIA's Vera Rubin platform, owned by the customer rather than the chip supplier.
By publishing board and rack designs under Open Compute Project, Meta is doing what it did with Open Rack a decade ago: set the standard, create a market, benefit from volume it didn't fund alone. If other operators adopt the Arm AGI CPU, component ecosystems scale, costs fall, and Meta benefits from economies it didn't have to underwrite.
Google Opens Ironwood TPU Training to Cloud Customers, Publishes Full Developer Stack
Google just made its most powerful custom silicon available for external training workloads.
Google Cloud published the full developer guide for training on Ironwood (TPU7x), its seventh-generation Tensor Processing Unit. The guide covers FP8 training recipes, kernel optimisation via the Tokamax library, SparseCore offloading for collective communication, and parallelism strategies across pods scaling to 9,216 chips. Each Ironwood chip delivers 4.6 PFLOPS of FP8 compute, 192GB of HBM3e memory at 7.4 TB/s bandwidth, and ICI networking at 9.6 Tb/s. A full superpod delivers 42.5 exaflops.
Why this matters:
NVIDIA Cloud Partners collectively passed one million GPUs at GTC this month, across hundreds of customers. Anthropic alone approaches that figure on TPUs.
Google's custom silicon programme now has a customer commitment that rivals NVIDIA's entire neocloud ecosystem in scale.
The emergent multi-vendor inference era we started tracking in Issue #98 with AWS/Cerebras and NVIDIA’s Groq 3 LPX has a third architecture in production. Google's Ironwood, NVIDIA's Vera Rubin + LPX, and AWS's Trainium + Cerebras each represent a different approach to the same problem. NVIDIA's ecosystem lock-in is real. Google's vertical integration from TPU to model to cloud is real. AWS's multi-vendor flexibility is real. The decision is now genuinely three-way for anyone choosing infrastructure at scale.
NextEra Gets Federal Approval to Build 10GW of Gas Power for Data Centres
America's largest utility company just got permission to build a lot of gas turbines.
NextEra Energy has received approval for 10GW of natural gas generation: 4.3GW in Pennsylvania ($17 billion) and 5.2GW in Texas ($16 billion). The company is targeting 15GW of new generation for data centre hubs by 2035 through its "15 by 35" origination channel, with current discussions covering 20 hubs.
Why this matters:
NextEra's "15 by 35" programme targets 15GW across 20 data centre hubs by 2035. That's more power than most countries' entire data centre markets.
The utility industry is building an entirely new business line around AI compute demand, indicating that data centre power is no longer a side contract but a core growth strategy.
Pennsylvania and Texas are deliberate choices. PJM Interconnection and ERCOT are the two largest power markets in the US and the two where data centre demand is growing fastest. NextEra is placing gas generation into the markets where the grid constraint is most acute and the willingness to approve fossil fuel generation is highest.
The Rundown
Two announcements this week captured where the industry is and where the questions live.
Musk said he'd build the world's most advanced chip factory. He has no timeline, no lithography machines, no semiconductor workforce, and a clean room philosophy that involves cigars. The demand signal underneath the announcement, however, is real: when a customer at that scale concludes the allocation queue won't serve them, the queue is the problem.
Whether Terafab produces chips or produces leverage, it changes the conversation between TSMC and its largest customers.
The FT published Nscale's predecessor defaulting on loans from a distressed credit investor at 17.5% interest, eighteen months before a $14.6 billion valuation.
The capital trajectory is real. The 8GW site is real. The NVIDIA relationship is real. The speed is real. The recent board hires are real. Yet the corporate history appears thinner than the headlines suggest.
Three publications in three weeks have now asked the same question: do the due diligence procedures and governance frameworks match the velocity?
The rest of the market kept building.
Meta added its fifth silicon partner. Google made Ironwood production-ready for external customers. NextEra got 10GW of gas power approved because the grid can't keep up. Elliott deployed billions into Synopsys, the EDA duopolist that designs every GPU on the market. And OpenAI killed Sora to chase enterprise revenue.
The ambition keeps scaling, but the scrutiny might finally be catching up.
See you next week.