Asymmetric Thesis: The Market Is Misreading The Threat Vector

The current noise around Google’s TPUs “disrupting” Nvidia and upending AI infrastructure economics looks dramatic on the screen, but I think it is structurally overstated and mis-angled. What is really happening is not a clean substitution of Nvidia GPUs by Google TPUs. It is a gradual diversification of hyperscaler supply and a repricing of power within the AI capital stack. That is a very different story for a portfolio concentrated in neoclouds and HPC miners than the recent drawdowns imply.

From a distance, the narrative is simple. Meta flirts with shifting part of its future capex to Google’s TPU roadmap, Google publicly positions TPUs as cheaper and more efficient, and Nvidia instantly sheds hundreds of billions in market value. Under the surface, however, the core infrastructure flywheels still run through Nvidia’s CUDA ecosystem, and independent GPU clouds like Nebius and HPC miners like Cipher remain the only scalable way for non-hyperscalers to access high-end compute at volume.

My contrarian take is the following. The circular financing loop between chip vendors and AI labs is indeed real, and it introduces systemic fragility. But it does not invalidate the neocloud/HPC thesis. On the contrary, it pushes value toward the players that can arbitrage three things at once: vendor concentration risk, capex cyclicality, and power pricing.

In that sense, TPUs are less an existential threat and more an additional input into the same circular system. For an AI infra-heavy portfolio, the real question is not “TPU versus GPU” but “who controls the bottleneck when the music briefly stops.”

The Illusion Of Substitution: TPUs As Edge Optimizers, Not Core Displacers

At the hardware level, Google’s TPUs are highly capable. TPU v5p is a dedicated matrix engine with around 95 GB of HBM3 and roughly 450 TFLOPS of bfloat16 compute per chip. Nvidia’s H100, by contrast, delivers 80 GB of HBM3 and up to about 4 PFLOPS of FP8 throughput, with the next-gen GB200 pairing a Grace CPU and Blackwell GPU to reach roughly 192 GB and close to 20 PFLOPS. On raw numbers alone, both sides can tell a compelling story.

But performance tables are not where power sits. Google’s TPUs live inside a tightly integrated software stack built around XLA and JAX, while Nvidia’s GPUs are embedded into CUDA, cuDNN, and a decade-plus of tooling and operator libraries that most serious AI teams treat as default infrastructure. Porting a non-trivial, end-to-end AI pipeline from CUDA-first to TPU-first is not just a compiler swap; it is an organizational change, a workflow change, and in many cases a staffing change. That is why, despite repeated launches of in-house accelerators by AWS, Microsoft, and Google, Nvidia still powers the overwhelming majority of visible AI training workloads and remains the neutral standard outside a small number of hyperscaler-controlled silos.

This is where the “TPUs are killing Nvidia” hot takes fall apart under scrutiny. Meta exploring multi-year TPU commitments is best viewed as a hedge against concentration risk and pricing power, not a religious conversion. Google itself is openly targeting perhaps 10% of Nvidia’s AI chip revenue over time, not 50% or 70%. Even after the Meta news, independent industry surveys still place Nvidia near 80% market share in AI accelerators, with Google TPUs and AMD GPUs each hovering around mid-single digits. In other words, TPUs are credible enough to pressure Nvidia’s pricing curve at the margin, but the ecosystem gravity still pulls workloads back to CUDA, especially outside Google Cloud.

For neoclouds and HPC data center miners, this asymmetry matters. They cannot buy TPUs. They rent and deploy Nvidia GPUs. As long as non-Google demand from OpenAI-style labs, enterprises, and sovereign clouds keeps scaling, the primary economic engine behind Nebius, Cipher, and IREN still runs on Nvidia’s roadmap, not Google’s. TPU progress can loosen supply tightness and compress GPU prices, but that acts more like an input cost shock than a demand shock for these names. The substitution story is tidy. The reality is messier, and far more favorable to specialized GPU clouds than the headline tape suggests.

Circular Capital Loops: When AI Labs, Hyperscalers, And Chip Vendors Finance Each Other

The more interesting risk is not TPU versus GPU at all. It is the circularity of capital flows between AI labs, hyperscalers, and chip suppliers. Nvidia and AMD invest in model labs and emerging AI clouds, those entities sign multi-year GPU spend commitments, and hyperscalers in turn lock in massive, often exclusive, contracts with the same chip vendors. The result is a loop where much of the capex being “spent” is effectively recycled within a tight ecosystem of counterparties.

The Meta–Google TPU discussions fit squarely into that loop. Meta reportedly contemplates billions in spend on Google’s TPUs, with an eye toward both cost efficiency and strategic leverage in its negotiations with Nvidia. At the same time, Nebius secures a $3 billion contract to provide Nvidia GPU clusters to Meta, and has already fully sold out existing capacity as AI training demand explodes. AI labs and platform companies are, in effect, intermediaries in a compute-for-equity and compute-for-contracts cycle that amplifies both booms and busts.

This circularity is not inherently fraudulent, but it does obscure true end-user demand. When Amazon commits up to $50 billion to AI supercomputing for US government customers, that is not purely organic, bottom-up software demand. It is a strategic bet on locking in infrastructure relevance that just happens to be booked as capex. Nvidia’s recent analyst pushback, emphasizing its GPUs as the “only platform running every AI model,” is partly a recognition that investors are beginning to question how much of its growth is driven by genuine, durable workloads versus a self-referential arms race among hyperscalers.

For a portfolio tilted toward neoclouds and HPC miners, the question is simple. Does the circularity live primarily at the lab–hyperscaler–chip vendor triangle, or does it also infect the independent GPU clouds and power-rich miners that are renting into that demand?

My read so far is that most of the circularity risk is upstream. Nebius, Cipher, and IREN are capacity providers. Their contracts may be long-dated and highly concentrated, which introduces tenant and renewal risk, but they are not paying equity into labs to create demand. They are selling megawatts and GPU hours into a frenzy created elsewhere. That distinction matters a lot when the cycle eventually normalizes.

Neoclouds In The Crossfire: Triangulating Nebius, Cipher, And IREN

Nebius offers the cleanest window into how this plays out operationally. The company has signed multi-billion dollar deals with Microsoft and Meta to provide Nvidia GPU clusters and posted more than a fourfold rise in revenue year over year, with management openly stating that current GPU capacity is fully sold out. On the surface, this looks like peak cycle behavior, and it probably is. Yet the economics are more nuanced. Nebius’ core bet is that supply scarcity today lets it entrench relationships and later extract better economics as the market matures and services move up the stack.

Cipher Mining represents a different angle on the same bottleneck. Originally a pure-play Bitcoin miner, Cipher has been pivoting part of its footprint into high-performance computing hosting, signing multi-year deals to lease AI-ready data center capacity, backed in some cases by large cloud and AI tenants. These contracts typically involve fixed or inflation-linked payments per megawatt with optionality on GPU density, which means Cipher is not as exposed to spot GPU pricing as a bare-metal neocloud. The flipside is tenant concentration and counterparty risk. If a flagship tenant scales back AI ambitions or renegotiates terms, CIFR’s contracted revenue becomes a point of fragility rather than strength.

IREN sits somewhere between the two models. Its roots are in Bitcoin mining, with a cost structure built around low-cost renewable power and self-owned infrastructure. As IREN expands into AI hosting, it can leverage that same power arbitrage and vertically integrated approach, but unlike Nebius it does not have the same depth of hyperscaler-grade GPU orchestration yet, and unlike Cipher it has a larger residual exposure to BTC price and mining economics. In practice, IREN can become a hybrid asset: partially correlated to Bitcoin cycles and partially leveraged to AI demand as it signs more HPC contracts. That hybrid profile makes its cash flows harder to model, but also creates more paths for upside if either leg outperforms.

When I triangulate across these three, a pattern emerges. Nebius is closest to a pure-play neocloud with maximal direct leverage to Nvidia’s GPU roadmap and hyperscaler AI budgets. Cipher is more of a long-dated landlord to the AI buildout, with heavy tenant risk but potentially more stable unit economics per megawatt. IREN is the optionality trade, where transition risk from BTC to AI is meaningful but so is the multi-cycle payoff if management executes. TPUs and capital circularity affect all three, but mostly indirectly through GPU pricing, tenant behavior, and future renewal terms rather than immediate business model disruption.

Where The Value Actually Accrues: Algorithms, Networks, And Incentives

To understand whether AI circularity should change your portfolio stance, you have to reverse-engineer incentives. Nvidia’s compensation and narrative are anchored in maintaining volume growth and preserving the premium of its full-stack platform. That means tolerating some TPU and custom-ASIC share loss if it can deepen CUDA’s grip on developers and keep its networking and software attach rates high. Google, by contrast, is not trying to become a neutral chip arms dealer. It wants to compress its own cost to serve AI workloads, differentiate Google Cloud, and gain negotiation leverage with Nvidia. Those goals are not mutually exclusive, and they do not automatically crush neoclouds.

Neoclouds like Nebius are incentivized to front-run demand by over-securing GPUs and power, even at thin margins, in order to build a network of customers and a reputation for being the “fast follower” alternative to the big three clouds. That creates execution risk if the cycle turns abruptly, but it also lays the groundwork for platform-like economics later. If Nebius can move from simply renting GPU hours to offering managed AI environments, pre-trained model libraries, and sovereign-cloud style compliance, the economics shift from capex recycling to recurring, higher-margin services. Cipher and IREN, in turn, are incentivized to use their balance sheets and low-cost power to enter long-dated hosting deals that derisk near-term cash flow while retaining optionality on re-pricing or re-tenanting over time.

From a network and algorithmic perspective, Nvidia still sits at the center. Its GPUs remain the default for almost every serious AI model outside proprietary in-house stacks, and four million-plus developers have already standardized their tooling on CUDA. That network effect is what allows Nebius to sell out capacity and Cipher to sign large AI hosting agreements. If TPUs were truly eroding Nvidia’s dominance, you would see a visible migration of third-party workloads toward Google Cloud and away from independent GPU providers. So far, we are seeing the opposite: Google’s TPUs are mostly consumed inside Google Cloud or via a few large partners, while everyone else continues to chase Nvidia capacity wherever they can find it.

When I plug this into the Multibagger Evaluation Protocol lens, the algorithmic and network scores for Nvidia and the best-positioned neoclouds remain high. The disruption risk from TPUs has ticked up, but mainly through its impact on pricing power rather than ecosystem collapse. That nuance is getting lost in the current volatility, and that disconnect, in my view, is where the asymmetry lies.

Actionable Takeaways: How I’d Position Around TPUs, GPUs, And Circularity

Pulling all of this together, I do not see Google’s TPU push or the visible AI circularity as thesis-breaking for an AI-infra tilted portfolio. I see it as an early stress test that reveals where the real fragilities lie. The most vulnerable actors are the highly leveraged labs and the hyperscalers that over-commit to one vendor and then need to use financial engineering or in-house chips to rebalance. The relatively under-discussed beneficiaries are the players that own power, land, and neutral GPU capacity and can underwrite those assets across multiple cycles.

If I translate that into a probabilistic stance, Nebius screens as the highest-quality neocloud in this specific context. It combines strong contracted demand from Meta and Microsoft with clear operational proof that its Nvidia-based clusters are fully utilized, even at thin margins. On my internal scoring, its organizational execution and algorithmic leverage are strong, its network position is improving, and its disruption risk from TPUs remains moderate. That yields a Multibagger Confidence score in the low seventies.

Cipher scores slightly lower on network and algorithmic leverage because its model is less software-rich, but it gains points on asset duration and cash-flow visibility through long-term hosting contracts, landing it in the high sixties. IREN, with its hybrid BTC plus AI profile, sits a touch below that today, with meaningful upside if it can scale AI hosting without losing its power-cost edge.

The key practical implication is straightforward. I would not rotate out of neoclouds or HPC miners purely because of TPU headlines or circularity discourse. Instead, I would tilt within the theme. Prefer operators with sovereign or multi-tenant exposure over single-tenant dependence, favor those that are already layering services on top of raw GPU time, and be willing to pay up for balance-sheet flexibility and diversified counterparties. In other words, the outlook for the portfolio does not fundamentally change, but the bar for capital allocation inside the AI infra bucket just got higher. That is a healthy adjustment, not a reason to abandon the space.