The Great TPU Unbundling: Google’s Quiet Push to Reshape the $150 Billion AI Chip War

Midjourney was burning $2.1 million a month on inference. Then it ripped out its NVIDIA clusters and plugged in Google’s TPU v6e chips. The monthly bill plunged to $700,000, a 65 percent drop. That sort of number gets attention—not just from the CFO, but from every hyperscaler and neocloud operator watching the tectonic plates shift under the AI infrastructure market.

But the Midjourney move wasn’t just a one-off cost story. It was the canary in the compute mine. Google had been quietly reshaping its TPU strategy for months, and by mid-2026, the rumblings had turned into an unmistakable roar: the search giant was unbundling its custom silicon and pushing it out the door to neoclouds—the independent GPU-rental firms that have become NVIDIA’s most fanatical customer base. The message was clear: TPUs are no longer just a Google Cloud managed service. They’re a product, and Google wants to sell them, finance them, and even rent them back from the people who buy them.

That pivot lands squarely in the middle of a $150 billion-plus AI chip market, expanding at a pace that has cloud providers writing capex checks that would have seemed insane two years ago. TrendForce expects the eight largest CSPs to spend over $710 billion in 2026, a 61 percent year-on-year leap. Alphabet alone is pegged for more than $178 billion, almost double its 2025 figure. Inside that spending spree, TPUs are set to power nearly 78 percent of Google’s own AI server shipments, making it the only hyperscaler where ASIC volumes actually outpace GPUs. So when Google starts treating those same ASICs as a merchant chip business, the NVIDIA stranglehold—86 percent of data center chip revenue, by some estimates—suddenly doesn’t look quite so permanent.

The Architects of the Unbundling

For years, TPUs were the quintessential walled garden technology. Designed in collaboration with Broadcom, manufactured by TSMC, and deployed entirely inside Google’s data centers, they were accessible only via Google Cloud’s rental model. That kept them out of the hands of anyone who wanted physical control of their hardware—banks, defense contractors, healthcare firms with sensitive data—and it meant Google never had to build a real go-to-market sales org for silicon. But the arrival of the eighth-generation TPU family, and the financial architecture Google has wrapped around it, marks a clean break.

The new playbook has three legs. First, direct outreach to neoclouds—the CoreWeaves and Nebiuses of the world—with a pitch that stresses not just raw performance per dollar, but a simpler, more stable networking stack than NVIDIA’s increasingly complex GPU-to-GPU fabrics. Second, a financing and lease-back model that turns Google into both chip seller and anchor tenant: it will help finance data center construction, provide the TPUs, and then lease back capacity for its own AI workloads, effectively underwriting the neocloud’s risk. Third, the Blackstone joint venture, announced in May 2026, which throws $50 billion in equity and $200 billion in project leverage behind a new AI company explicitly built to compete with NVIDIA-backed neoclouds. That venture alone plans 500 megawatts of TPU capacity by 2027, with Google supplying the iron and Blackstone handling the capital stack.

Insiders have told The Information that Google’s internal target is to reach roughly 10 percent of NVIDIA’s AI server chip revenue. That might sound modest, but in a market headed toward $300 billion a year, 10 percent is tens of billions of dollars—and more importantly, it gives Google the scale to negotiate harder on TSMC’s advanced packaging lines, where NVIDIA currently calls the shots. Morgan Stanley analysts estimate that just half a million TPUs sold externally could generate around $130 billion in revenue for Google by 2027. The math is brutal and simple: sell more chips, lower the unit cost, fund the next generation, and pull a larger share of the value chain away from Santa Clara.

The Silicon: From Trillium to Sunfish

Underpinning all this is a product line that has evolved from a narrow matrix-multiplication engine into a bifurcated, workload-specific portfolio. Trillium, or v6e, is the current backbone. In an 8-chip configuration it delivers 7,344 TFLOPS with 256 GB of HBM, roughly matching a quad-H100 NVL setup while drawing 300 watts per chip versus the H100’s 700. With committed-use discounts on Google Cloud, v6e pricing drops to $0.39 per chip-hour, and in real-world large language model training, Google claims it can deliver up to 4x better price-performance than H100s. That’s the kind of delta that made Anthropic commit to hundreds of thousands of Trillium chips in November 2025, scaling toward a million by 2027.

Then came Ironwood, the seventh generation, built on TSMC’s 3nm process. At 4.6 petaFLOPS of dense FP8 per chip, it barely edges out NVIDIA’s B200, and when clustered into a 9,216-chip pod, it hits 42.5 exaFLOPS—more than 24 times the peak performance of the world’s then-fastest supercomputer, at least in the FP8 precision Google uses for AI. Ironwood packs 192 GB of HBM per chip, a 6x leap over Trillium, and Google makes a point of its carbon efficiency: a 3.7x improvement in Compute Carbon Intensity over v5p, and fleet-wide measurements showing a 5x jump in utilized FLOPs per watt.

On Hacker News, the benchmarks sparked both applause and skepticism. “Per chip, TPU 7x has 192GB of HBM3e, whereas the NVIDIA B200 has 186GB,” one commenter noted, a data point that had engineers recalculating memory-bandwidth economics. Another, however, pointed to the split between training and inference chips that Google formalized with the eighth generation: “Splitting TPUs into dedicated training vs inference chips feels like an admission that the bottleneck has shifted from FLOPs to memory bandwidth + latency.” That split became official at Google Cloud Next 2026, where the company unveiled TPU 8t (“Sunfish”) for training and TPU 8i (“Zebrafish”) for inference. Sunfish targets frontier model training with 12.6 peak FP4 petaFLOPS and a new SparseCore accelerator for embedding lookups, while Zebrafish focuses on agentic workloads with 288 GB HBM, 384 MB on-chip SRAM, and a “Boardfly” architecture that cuts maximum network diameter by more than half. Google says Zebrafish delivers 80 percent better performance-per-dollar than the previous generation for large mixture-of-experts models.

The community took notice. “Virtually all the frontier AI labs use TPU. The only one that doesn’t use TPU is OpenAI due to the exclusive deal with Microsoft. Given the newly launched Gen 8 TPU this month, it’s likely OpenAI will contemplate using TPU too,” a Hacker News user wrote. That comment may be more wish than reality for now, but it captures a growing sentiment: when even the most CUDA-locked labs start hedging, the ecosystem is changing.

NVIDIA’s Countermoves

NVIDIA didn’t sit still. Almost as soon as Google’s neocloud outreach became known, the GPU giant rolled out a lease-back guarantee of its own. Under that program, if a neocloud partner can’t rent out its next-gen Vera Rubin GPUs, NVIDIA will buy the capacity back at roughly 50 percent of the expected market rental price; if the partner earns more, NVIDIA takes half the excess. It’s less generous than a standard market lease, but it creates a revenue floor that helps neoclouds secure bank financing—exactly the kind of sweetener that keeps them in the NVIDIA camp.

Then there was the Nscale episode. Nscale, a two-year-old neocloud in which NVIDIA is a major investor and preferred shareholder, was reportedly one of Google’s early targets. According to sources, NVIDIA discussed new financial incentives with Nscale to keep it away from TPUs. Nscale’s spokesperson denied any such quid pro quo, stating that all signed clusters remain GPU-backed. But the fact that NVIDIA felt the need to offer anything at all is telling. Neoclouds, after all, are where the next wave of AI capacity is being built; CoreWeave alone had locked in $99.4 billion in backlog orders, and NVIDIA had sunk another $2 billion into its stock in January 2026. Losing even a fraction of that channel to Google would be a strategic headache.

Broader Dynamics and Supply Chain Tremors

Google is hardly alone in trying to pry open the AI chip market. Amazon’s Trainium3 is ramping into production, Microsoft has its Maia 200 for inference, and Meta’s MTIA program is pushing forward, despite reports of software-tuning delays. AMD’s Instinct MI400 series, built on TSMC’s 2nm node, is expected to grab 12-15 percent of the AI accelerator market by Q4 2026, with the MI455X targeting memory-intensive workloads that have historically been NVIDIA’s sweet spot. Every large CSP is now pursuing an ASIC strategy, but only Google has started selling its ASICs to third parties. That move creates a tiered market where custom chips compete directly with GPUs, and it injects a new variable into the supply-demand balance: if neoclouds can get TPUs on better terms than GPUs, NVIDIA’s pricing power starts to erode.

That possibility is already being weaponized in negotiations. Industry sources say that clients now routinely use TPU availability as a bargaining chip: one large lab, reportedly OpenAI, secured a 30 percent discount on GPU purchases simply by threatening to switch. “You can’t just say ‘we can’t get GPUs’ anymore,” one procurement executive told SemiAnalysis. “The customer will ask, ‘what about TPUs?’” Even if no machines actually change hands, the threat of substitution is deflating NVIDIA’s historically untouchable 75 percent gross margins.

But the unbundling has its own supply chain fault lines. Google’s relationship with Broadcom, its long-time TPU design partner, is under strain. Broadcom gets up to 70 percent gross margins on TPU work, and Morgan Stanley estimates that TPU-related revenue could hit $80 billion for Broadcom in fiscal 2027. Yet Google has openly courted MediaTek for future TPU designs, with a credible 3nm project already underway. The motivation is straightforward: MediaTek is willing to accept margins closer to 30 percent, dramatically lowering Google’s cost basis. A July 2026 Morgan Stanley note argued that Broadcom will likely retain around 80 percent of TPU market share long-term, with MediaTek capturing 15-20 percent, but the mere threat of a second source has weighed on Broadcom’s stock. The broader impact? As TPU volumes scale to millions of units, the ability to play suppliers off each other becomes essential to keeping unit economics viable.

Developer Reality Check

For all the hardware progress, the software story remains TPU’s biggest vulnerability—and the source of the most venom in developer forums. On Google’s own developer boards, a thread from May 2026 asked, “Anyone here actually using TPUs in production workloads yet? Most AI infrastructure conversations still end up centered around GPUs.” The answers were mixed: some shared success stories with JAX and vLLM, while others complained about the free tier “taking 40 seconds just to execute import pandas as pd.” GitHub issues reveal deeper friction: “Jax can’t use a TPU’s CPU, and… there are no plans to add this to Jax,” one developer noted, pointing to a fundamental architectural limitation that complicates certain workloads.

Midjourney’s CEO has publicly vented that the company’s research progress fell a year behind because of TPU ecosystem shortcomings. The core problem is that rapid experimentation still belongs to GPUs and PyTorch. “For a new idea, it takes a few hours on GPU; on TPU, just setting up the environment takes days,” one ML researcher summarized. Google’s fixes—TorchTPU, a collaboration with Meta to get PyTorch running natively on TPU, and tighter integration with vLLM—are promising but still young. In July 2026, Google published a technical report showing how it optimized Qwen 3.5-397B MoE on Ironwood to achieve 3.1x improvement for decode-heavy workloads, but such results require deep TPU-specific tuning that most teams can’t afford.

Nevertheless, some large adopters have found workable strategies. Anthropic’s three-platform approach—NVIDIA GPUs for research, TPUs and Amazon Trainium for large-scale training and inference—has become a reference architecture. The tradeoff is tripled operational complexity, but the payoff is a TCO that can be 50 percent lower than a pure NVIDIA deployment, plus real supplier independence. “Study on GPU, scale on ASIC” is evolving into a playbook that even moderate-sized AI firms can consider, particularly as inference workloads come to dominate.

What This Means for the $150B Battlefield

Google’s TPU unbundling is not just a product launch; it’s a structural bet on how AI infrastructure will be consumed over the next decade. If training is increasingly dominated by a handful of mega-labs and inference becomes the volume business—projected to consume 75 percent of AI compute by 2030—then the market tilts toward chips optimized for cost-effective inference at scale. That is precisely where TPU 8i, with its massive on-chip SRAM and agent-focused architecture, is aimed. NVIDIA, for its part, is not standing still. Its lease-back guarantees, its deepening financial ties to neoclouds, and its relentless cadence of new architectures keep its ecosystem sticky. But the emperor’s clothes have a few holes now.

On Hacker News, the debate underscores the uncertainty: “Safe bet Google’s TPUv5 will be competitive with the H100, as the v4 was with the A100, but their offering clearly hasn’t impacted market share thus far and there is no indication Google intends to make their chips available outside of GCP.” That comment, written before the Blackstone deal and the neocloud outreach, now reads like a relic. Google is making its chips available outside of GCP, and it’s doing so with billions in financial backing. Whether that changes market share depends on execution—on how fast TorchTPU matures, on whether TSMC can pump out enough advanced packaging to supply both NVIDIA and Google, and on how aggressively NVIDIA defends its turf. But for the first time, the question isn’t if TPUs will become a merchant market, but how large that market can get.

In the end, the most powerful thing Google may have done is simply offer an alternative. The mere existence of a credible second source changes every conversation in the AI infrastructure supply chain. When a startup can walk into a meeting and say, “We’re considering TPUs,” the pricing power slides, the incumbency advantage shrinks, and the $150 billion prize gets redistributed. That redistribution has already begun.