Yesterday we connected two stories — Apple’s rumoured M7 chip pivot and Nvidia’s Kyber rack delay — and noted they converge on the same 2027-2028 timeline. Today a new piece of the puzzle emerges: Apple’s reported “Baltra” AI server chip, developed with Broadcom, is targeting 2027 server deployment. That means Apple isn’t just betting on on-device AI. It’s simultaneously building server-class silicon. And both projects may be drawing from the same finite manufacturing capacity that Nvidia needs for Rubin Ultra.

Let’s be explicit about what we’re doing: we’re reading publicly reported supply chain analysis and connecting dots. We have no inside knowledge of TSMC’s allocation decisions, Apple’s manufacturing bookings, or Nvidia’s production schedules. What follows is informed speculation grounded in public reporting — the same kind of dot-connecting we did yesterday, now with a deeper supply chain lens.

🔍 THE BOTTOM LINE

The Kyber delay’s root cause may be deeper than a single unmanufacturable PCB midplane. The entire advanced semiconductor manufacturing stack — TSMC wafer capacity, advanced packaging lines, and ASML’s EUV lithography tools — is running at 100% utilisation with multiple companies fighting for the same finite resources. Apple’s simultaneous M7 transition and Baltra server chip project could be squeezing the same manufacturing lines Nvidia needs, making the Kyber delay a symptom of a structural capacity war rather than just a PCB engineering problem.

Three Rounds

This isn’t the first time Apple and Nvidia have squared off. It’s the third.

Round 1 — Mobile: A decade ago, Apple’s A-series chips and Nvidia’s Tegra fought for the mobile frontier. Apple won decisively. Nvidia exited mobile entirely, pivoting to GPUs and AI accelerators.

Round 2 — Pro Computing: Apple’s M-series transition dropped Nvidia GPUs from the Mac lineup. Apple built its own GPU architecture into its silicon, cutting Nvidia out of the premium creative workstation market. Nvidia didn’t lose — it just stopped being invited.

Round 3 — AI Infrastructure: As one titan stumbles on AI manufacturing — Nvidia’s Kyber rack delayed to 2028 over a PCB midplane — the other is taking advantage. Apple’s reported Baltra server chip and M7 transition are booking the same TSMC advanced packaging and wafer capacity Nvidia needs for Rubin Ultra. The fight has moved from who makes the best chip to who can secure enough manufacturing capacity to build any chips at all.

But here’s the tension that makes this round different from the first two: Apple isn’t winning. Apple is stumbling too.

The Titan That Wasn’t Ready

Let’s not pretend Apple saw the AI arms race coming. By every public indicator, Apple was caught flat-footed by the speed of the AI explosion. Siri lagged. Apple Intelligence arrived late and underwhelming. The company that proudly announced “AI for the rest of us” spent 2024-2025 watching OpenAI, Google, and Anthropic define the frontier while its own AI features were delayed, scaled back, or quietly forgotten.

The reported M6 skip isn’t a masterstroke of strategic foresight. It’s a panic pivot. You don’t skip an entire chip generation — leaving the premium Mac lineup to rely on an M5 Ultra stopgap — if you’re executing a long-term plan. You do it if you looked at the road you were on, realised it wouldn’t get you to AI-capable in time, and decided to burn the roadmap and start over. That’s not confidence. That’s urgency.

The Baltra server chip project tells the same story from the other direction. Apple built its Private Cloud Compute infrastructure on existing silicon — essentially repackaged M-series chips in server racks. Baltra, reportedly purpose-built for AI inference, suggests Apple realised that wasn’t enough. The general-purpose approach didn’t scale. They needed a dedicated AI server chip, and they needed it by 2027.

So the framing isn’t “Apple the strategic genius exploiting Nvidia’s weakness.” It’s more honest than that: two titans are stumbling. Nvidia stumbled first — its rack-scale manufacturing hit a wall. Apple stumbled earlier — it missed the AI turn entirely and is now racing to catch up. What’s new is that Apple’s scramble to catch up happens to require the same manufacturing resources Nvidia needs to fix its own stumble. Neither company is in control. Both are reacting. The supply chain is the referee.

The pattern across all three rounds isn’t that Apple always wins. It’s that Apple changes the game — sometimes because it’s ahead, and sometimes because it has no other choice.

The Baltra Piece

According to TrendForce reporting, Apple is pushing an in-house AI server chip codenamed “Baltra” for 2027 server deployment. The chip is reportedly designed primarily for AI inference — running models, not training them. Wccftech and TechPowerUp have also reported on Baltra’s inference-focused design, with manufacturing reportedly handled by TSMC.

Separately, Apple and Broadcom expanded their custom chip supply partnership through 2031 — a deal that covers both connectivity components and, reportedly, processor design collaboration. Broadcom is also working with OpenAI on its own AI chips, making it a central player in the “build your own AI silicon” movement.

The Baltra project adds a third leg to Apple’s AI silicon strategy we outlined yesterday:

• M7 (Andros) → on-device AI for consumers (2027-2028)
• Baltra → server AI chips for Apple’s own cloud inference infrastructure (2027)
• Both converge on the same manufacturing window as Nvidia’s delayed Kyber rack

Apple isn’t choosing between scale-up and scale-down. It’s doing both. And both need the same manufacturing resources.

The TSMC Capacity Squeeze

Here’s where the supply chain picture gets interesting — and where we move from reporting into speculation.

According to SemiAnalysis, Apple and Nvidia have historically used different TSMC packaging technologies. Apple used InFO (integrated fan-out) for its A and M series. Nvidia used CoWoS (chip-on-wafer-on-substrate) for its GPUs. They operated in separate lanes.

That separation is dissolving. Igor’s Lab reported in January 2026 that Apple’s M5 Pro and M5 Max are transitioning to SoIC-MH packaging — a technology that allows horizontal and vertical stacking of multiple chips. This puts Apple into the same AP6 and AP7 TSMC packaging lines that Nvidia uses for its CoWoS-based AI accelerators.

The result, per Igor’s Lab: “This convergence increases the risk of bottlenecks in advanced 3D packaging, especially against the backdrop of the rapidly growing AI market.”

Meanwhile, Digitimes reported that TSMC has secured large advanced packaging orders from both Apple and Nvidia simultaneously. And TweakTown reported that TSMC’s 3nm and 5nm capacity is expected to be “100% booked” in 2026 — with Apple and Nvidia together accounting for over 40% of TSMC’s revenue.

If Apple is pulling capacity for both the M7 consumer transition and the Baltra server chip project, that’s a massive draw on the same foundry Nvidia needs for Rubin Ultra. The question isn’t whether there’s enough demand — it’s whether there’s enough manufacturing.

The ASML Bottleneck — One Company, One Tool

The squeeze goes deeper than TSMC. The most advanced chips in the world — the 3nm, 2nm, and beyond — require EUV (extreme ultraviolet) lithography machines. There is exactly one company that makes them: ASML, headquartered in the Netherlands.

ASML is the only company on Earth capable of manufacturing EUV lithography systems. No Chinese, American, Japanese, or Korean company has a competing product. The machines cost roughly $400 million each and take years to build. TechPowerUp reported that ASML plans to deliver only 56 Low-NA and 10 High-NA EUV tools in all of 2027.

Those 66 machines have to serve TSMC, Intel, Samsung, and SK hynix — every major foundry simultaneously. If Apple’s M7 and Baltra chips require the most advanced process nodes (3nm or below), and Nvidia’s Rubin Ultra also needs those nodes, they’re all queueing for the same ASML-equipped TSMC lines.

This is not a market with a quick fix. You cannot build a new EUV foundry in 18 months. The constraint is physical — it’s the number of machines that exist, the rate at which ASML can produce them, and the years it takes to install and calibrate each one.

The PCB Layer — A Symptom, Not the Root Cause?

The SemiAnalysis report on Kyber’s delay pointed to a specific component: the PCB midplane — a multi-layer printed circuit board that connects every GPU in the rack. Manufacturing it at the density and layer count Nvidia’s design requires is “beyond what current PCB fabrication can reliably produce at volume.”

But TrendForce has reported that “Rubin’s cableless architecture and ASIC high-layer HDI designs are pushing PCBs to the center of AI compute power” — meaning the PCB complexity is an industry-wide escalation, not a one-off engineering miss. The key suppliers of high-layer-count HDI substrates are Japanese (Ibiden, Shinko Electric) and Taiwanese (Unimicron). These are the same suppliers that serve the broader advanced packaging market.

Here’s the speculative leap: if TSMC’s advanced packaging capacity is being split between Apple’s M7/Baltra projects and Nvidia’s Rubin Ultra, and if the PCB substrate suppliers are also seeing surging demand from multiple chipmakers simultaneously, then the Kyber delay might not be an isolated manufacturing failure. It might be the visible symptom of a capacity war where every advanced manufacturing resource — wafer fabrication, advanced packaging, PCB substrates, EUV lithography — is running at or beyond its limit.

Nvidia’s PCB midplane problem may be real, but the question we’re asking is whether it’s the root cause or just the place where the capacity crunch became visible. When every supplier in the chain is overbooked, delays don’t appear as “we couldn’t get factory time.” They appear as “this specific component is hard to make.” The framing changes the story.

The Monopoly Map

The advanced semiconductor manufacturing stack is stunningly concentrated. The EUV lithography machines that make 3nm and 2nm chips possible come from exactly one company — ASML, headquartered in the Netherlands. Not Sweden, not the US, not Japan. One Dutch company, one machine, no competitor on Earth.

The relevant monopolies and near-monopolies in the advanced manufacturing stack are:

The entire advanced AI chip supply chain runs through roughly half a dozen companies in three countries. When Apple and Nvidia both need the same TSMC lines, the same ASML tools, and the same PCB substrate suppliers, “capacity” becomes a zero-sum allocation problem.

The Production Shift Hypothesis

Could the processor shortages already relate to production being shifted? It’s the sharpest version of this theory. If Apple is transitioning its entire premium Mac lineup (M6 → M7) and launching a new server chip line (Baltra) simultaneously, the manufacturing capacity pre-allocation for those projects would have been locked in 12-24 months ago. TSMC doesn’t do last-minute capacity allocation for 3nm wafers — the lines are booked years in advance.

That means the capacity Apple is using for M7 and Baltra today was likely reserved during a window when Nvidia was also trying to lock in Rubin Ultra production. If TSMC had to choose between Apple’s massive consumer+server order and Nvidia’s rack-scale GPU order for the same AP6/AP7 packaging lines, the allocation decision itself could have contributed to the manufacturing timeline that made Kyber’s PCB midplane “unmanufacturable at scale.”

This is speculation. We cannot verify TSMC’s internal allocation decisions. But the structural conditions for it are real: finite capacity, multiple massive customers, and years-long lead times for the most advanced manufacturing resources on Earth.

NZ Angle

New Zealand has no domestic semiconductor manufacturing and no seat at the table in this capacity war. But the downstream effects are direct: every AI workload run from NZ depends on chips manufactured through this constrained supply chain. If Apple’s Baltra and M7 transition is squeezing Nvidia’s Rubin Ultra timeline, the cost per token of running frontier AI models in 2027 stays higher for longer — and NZ researchers, startups, and universities renting cloud GPUs pay that premium.

The capacity war also explains why NZ’s path to AI capability can’t rely on renting access to other people’s chips. The countries and companies that control manufacturing allocation will always prioritise their own AI infrastructure first. Aotearoa’s sovereign AI argument — the one we’ve made before — gets stronger when you realise the supply chain has a physical ceiling, and the ceiling is made by one Dutch company and one Taiwanese foundry.

❓ FAQ

Is Apple’s Baltra chip confirmed? No. Baltra has been reported by TrendForce, Wccftech, and TechPowerUp based on supply chain sources. Apple has not officially announced it. Everything in this article about Baltra is based on publicly reported rumours and supply chain analysis.

Could Apple’s capacity booking actually cause Nvidia’s delay? We’re speculating. The structural conditions exist — same TSMC lines, same packaging technology, finite capacity. But we cannot verify TSMC’s internal allocation decisions. The Kyber delay’s immediate cause is the PCB midplane. Whether capacity pressure contributed to that specific failure is a question we’re asking, not answering.

Why is ASML a monopoly? EUV lithography took 20+ years and tens of billions of euros to develop. ASML is the only company that succeeded. No competitor has a working EUV machine. The technology is so complex that the US, China, Japan, and Europe have all tried to build alternatives — none have succeeded at the leading edge.

🔍 THE BOTTOM LINE

Yesterday we saw two companies converging on the same timeline. Today we see why: they’re also converging on the same manufacturing resources. The AI arms race isn’t just about who designs the best chip — it’s about who can book enough capacity at the one foundry, the one lithography company, and the handful of substrate suppliers that make the whole thing possible. Apple’s Baltra project and M7 transition may be pulling capacity from the same lines Nvidia needs. If so, the Kyber delay isn’t just a PCB problem. It’s the first visible crack in a supply chain running at absolute capacity.

📰 Sources

• SemiAnalysis (Apple-TSMC partnership analysis, Kyber delay reporting)
• Igor’s Lab (Apple and Nvidia TSMC advanced packaging convergence, January 2026)
• TrendForce (Baltra server chip reporting, PCB substrate analysis)
• Digitimes (TSMC advanced packaging orders from Apple and Nvidia)
• TechPowerUp (ASML EUV tool delivery projections)
• Reuters (Nvidia advanced packaging comments)