Executive Summary
- The training phase monopoly is coming to an end. Scarcity and speed were key factors in Nvidia’s dominance from 2022 to 2024. The governing metric changes from time-to-solution to cost-per-token as the workload moves to inference.
- Nvidia’s unit economics are being completely destroyed by Google’s vertical integration. About 20% of Nvidia’s cost structure is used by the TPU v6e Trillium. An efficiency advantage of four to six times is produced by this internal margin capture and system level optimisation.
- The CUDA platform’s protective moat is quickly eroding. Software and underlying hardware have been successfully separated by PyTorch 2.0, JAX, Triton, and MLIR. Proprietary GPUs are becoming commodities as a result of this development.
- A significant amount of institutional capital is starting to disappear. Buffett’s 4.3 billion dollar Google stake, Druckenmiller’s positions, and Bridgewater’s 65% decrease in Nvidia exposure all point to a strong conviction. They think the capital cycle is turning around and that the valuation has peaked.
- There is a lot of risk associated with terminal valuation. A multiple contraction is predicted by concentration, where four direct hyperscaler competitors account for 61% of revenue, along with margin compression. Regardless of future revenue growth, a structural de-rating from forty times to twenty times sales is imminent.
The Great Bifurcation
The market’s prevailing belief that Nvidia Corporation is the unchallenged sovereign of the artificial intelligence revolution, perpetually levying a toll on every global floating-point operation, suffers from a fatal structural flaw. This story was true between 2022 and 2024, when the generative AI cycle was in its training phase. But it completely ignores the geometric and economic changes that will characterise the upcoming inference and deployment phase, which is anticipated to occur between 2025 and 2030.
We are observing the initial destabilisation of a Great Bifurcation in the AI compute market.
For the past eighteen months, the “Nvidia Long” trade has been the most popular institutional position. It is based on the unwavering reliance on hyperscalers and the certainty of endless margin preservation. This fundamental premise is quickly approaching a structural breakdown point, according to our in-depth analysis.
The critical error embedded in the consensus view is the conflation of two fundamentally different economic regimes.
The 2022–2024 cycle was characterised by the Training Phase, a market driven by scarcity where cost was secondary and time-to-solution was the only important metric. To finish the model first during training, you are willing to pay any price.
Nvidia’s Blackwell architecture is the best in this situation.
However, the economic logic completely reverses as foundation models such as Gemini, GPT-4, and Llama 3 become widely deployed. Cost-per-token and tokens-per-watt regulate the Inference Phase, a utility game.
At Bancara, we advise caution when it comes to the widespread expectation that merchant silicon will maintain a constant margin. Nvidia’s pricing power is under significant deflationary pressure due to the combination of Alphabet’s Trillium TPU v6 architecture, the maturity of the JAX and XLA software pipeline, and the aggressive unit economics of the Gemini ecosystem.
This combination directly challenges the company’s long-term intrinsic value.
Google does more than just compete.
It is actively developing a post-GPU paradigm that puts overall cost of ownership and system throughput ahead of isolated device specifications. Google creates an AI inference cost structure that merchant silicon vendors cannot mathematically match by cutting off its reliance on the “Nvidia Tax”, a markup estimated at 400–500% over the cost of the silicon.
The capital allocation ramifications are enormous.
The significant cost difference between using merchant GPUs and custom TPUs will require a thorough reevaluation of the entire AI hardware industry when inference workloads surpass training workloads in late 2025.
The Silicon Geometry: Trillium vs. Blackwell
One must look beyond straightforward comparisons of individual chip specifications, which are frequently found in standard notes, in order to understand the risk Google poses to Nvidia. The architectural philosophy is where the main distinction lies.
Both companies’ financial paths will be determined by this differentiation.
Scale-Up vs. Scale-Out: The Philosophical Divide
Nvidia’s Blackwell architecture epitomises the pinnacle of scale up design.
In order to support general purpose workloads across a wide range of clients, it incorporates enormous compute density, substantial thermal output, and significant complexity into single nodes, such as HGX or DGX.
The Blackwell B200 is designed to be the all-purpose supercomputer. Its sheer computational power allows it to handle any task, from large language model training to meteorological modeling. This chip is essentially designed to perform well on its own.
On the other hand, the scale-out approach is exemplified by Google’s Tensor Processing Unit architecture, particularly the sixth-generation Trillium (v6e) and its Ironwood implementation. It views each chip as a single tile in a huge, optically connected mosaic rather than as the center of the universe.
Google’s design philosophy optimises the interconnects between thousands of moderately powerful chips to operate as a single, massive supercomputer, purposefully sacrificing single-device peak performance for system-level coherence.
This distinction is absolutely vital.
FLOPS or raw processing power are rarely the limiting factor for modern Large Language Models. Rather, memory bandwidth and interconnect latency are the real constraints. When operating at the scale of systems like Gemini or GPT-4, a scale-out deployment strategy is required due to the intrinsic structure of transformer models, which requires extensive communication across parameters.
The Trillium Threat Vector
The Trillium generation represents a tipping point in this architectural war because it closes the usability gap while widening the efficiency gap.
According to Google, Trillium offers over a 4x improvement in training performance, a 3x increase in inference throughput, and a 67% increase in energy efficiency when compared to the TPU v5e. These are generational leaps that change the data center build-out calculus, not incremental improvements.
In terms of raw specs, the Nvidia B200 continues to rule. It has a significant advantage over the TPU v6e, outperforming it by 300 to 400 percent in some raw compute benchmarks, particularly in FP8 compute. For the economics of inference workloads, this benefit is becoming more and more insignificant.
A Formula 1 engine is comparable to the B200. Although it provides unmatched performance, its operation requires a significant amount of energy, money, and maintenance.
The electric freight train is the TPU v6e. It is designed to minimise the cost per watt while transporting large token loads.
Optical Circuit Switching: The Hidden Advantage
Google’s secret weapon is Optical Circuit Switching (OCS) technology, which eliminates the need for power-hungry electrical transceivers at each hop by enabling Google to instantly change the topology of its supercomputers. Importantly, this optical interconnect fabric allows the Ironwood TPU pods to scale to thousands of chips, achieving multi-exaflop aggregate performance.
Because of this architectural decision, Google has a structural energy advantage that Nvidia finds difficult to match due to the typical data center infrastructure needs of its wide range of clients. Google’s ICI (Inter-Chip Interconnect) with OCS enables less expensive, lower-power scaling to 10,000+ chips, whereas Nvidia uses copper-based NVLink interconnects with high power consumption.
This is the distinction between selling a supercomputer as a whole and selling a chip.
The Axion Front: Full-Stack Control
Both accelerators and the host processors in charge of data management compete for silicon supremacy.
Nvidia created the Grace Hopper “Superchip” ecosystem in an attempt to dominate this layer with its Grace CPU. But Google’s launch of the Axion processor, its top-tier custom Arm-based CPU for data centers, offers a strong defense against Nvidia’s ambitious plans.
Axion provides exceptional performance and power efficiency by utilising the Arm Neoverse V2 platform. According to performance metrics, Axion instances (C4A) perform about 10 percent better than similar AWS Graviton4 instances. Google optimises the server rack’s entire thermal and data envelope by controlling both the CPU, Axion, and the Accelerator, TPU. It is impossible to achieve this level of optimisation in a setting that depends on several vendors.
The Economic War: Deconstructing the “Nvidia Tax”
The most immediate and quantifiable threat to Nvidia’s stock price is the exposure of the gross margin disparity between merchant and custom silicon. According to industry research, Google gets its AI processing power for about 20% of what rivals pay for expensive Nvidia GPUs. This suggests that Google has a 4x–6x cost-efficiency advantage per unit of compute.
The Anatomy of the Tax
Purchasing an H100 or B200 from a vendor such as Microsoft Meta or a sovereign wealth fund entails paying for Nvidia’s significant R&D amortisation, its 75% gross margin, the TSMC manufacturing premium, and the proprietary CUDA ecosystem fee. Google’s internal deployment of a TPU, on the other hand, only results in direct manufacturing costs to TSMC and the amortisation of its own R&D. The expensive intermediary rent extraction is effectively eliminated by this structure.
This cost advantage manifests directly in the pricing power of Google’s cloud services. The comparative unit economics are stark:
- TPU v5e Inference Cost: Inference costs are as low as $0.30 per million tokens thanks to Google’s JetStream optimisation on TPU v5e.
- H100 Cloud Rental: The market rate for renting H100s ranges from $2.85 to $4.50 per hour.
- Efficiency Delta: Analyses indicate that TPUs can be 4x more cost-effective per unit of compute than H100s for specific large-scale inference tasks.
After switching to TPUs, businesses like Midjourney have reported a significant 65% decrease in inference costs.
This improvement is not insignificant. New business models are made possible by this fundamental change.
For example, if a startup uses H100s to achieve 40% gross margins on an AI wrapper, switching to TPUs could significantly increase those margins to 70% or 80%.
The Inference Trap
When inference volume exceeds training volume in late 2025, Bancara’s internal models predict a structural ceiling on merchant chip margins. We refer to this circumstance as the “Inference Trap”, which presents Nvidia with a strategic conundrum. In order to compete with TPUs’ internal cost efficiency, the company must either reduce its high gross margins, which are currently around 75%, which will erode earnings, or give up the high-volume market segment to proprietary custom silicon.
The mathematical reality is unforgiving.
The Total Addressable Market for AI silicon clearly moves in the direction of the most economical architecture as the volume of inference rises sharply. Nvidia will lose the utility-scale inference market if it is unable to compete on price. Its earnings per share growth and, consequently, the value of its stock will plummet if it competes.
The Circular CapEx Illusion
Currently, a large portion of Nvidia’s income comes from what analysts refer to as “circular financing” or the arbitrage of cloud rentals. Hyperscalers purchase GPUs and lease them to AI startups, many of which they have invested in. These startups then use venture capital to cover the cost of the cloud. This leads to unstable revenue quality that is reliant on the AI funding cycle continuing.
The current market dynamic is totally upended by Google’s strategic shift.
By using TPUs for its extensive internal operations, such as Waymo, YouTube Gemini, and Search, Google drastically lowers the total addressable market that merchant silicon providers can access. Nvidia effectively loses control over the most valuable workload of the artificial intelligence era if Google Search, arguably the largest AI application on the planet, runs solely on TPUs.
The End of CUDA Lock-in
The CUDA Moat has been the main defensive argument used by Nvidia bulls for almost twenty years. According to the argument, switching to alternative hardware is either technically or financially impractical due to Nvidia’s proprietary software stack’s superior performance and deep integration into the developer ecosystem.
This view, while historically valid, is rapidly becoming obsolete in the face of open standards and new compiler technologies.
PyTorch 2.0 and Hardware Agnosticism
PyTorch 2.0 represents a watershed moment in AI software.
It presents torch.compile, a feature that creates effective kernels for various hardware targets by using the TorchInductor backend. The architecture is intended to be backend-agnostic, even though it currently produces Triton kernels for Nvidia GPUs. This significantly lowers the friction involved in transferring a model from a GPU to a TPU or to an AWS Trainium chip.
Researchers’ high-level Python logic is becoming more and more detached from the underlying hardware.
Nvidia’s market dominance is seriously threatened by this split.
The clear benefit of customised CUDA kernels fades as compilers get better at optimising code without regard to hardware. For the vast majority of commercial deployments, the adequate performance provided by automated compilers will be more than sufficient, even though CUDA may still be useful for extracting the absolute peak performance.
This is particularly true when considering the quadruple cost savings on the hardware itself.
JAX: The Native Tongue of the TPU
Google’s JAX framework is not merely a competitor to PyTorch.
It represents a paradigm shift in numerical computing.
This architecture aligns perfectly with the TPU design.
In the past, PyTorch relied on the Multiple Program Multiple Data imperative eager execution model. The functional programming model used by JAX is excellent for Single Program Multiple Data.
There are significant economic ramifications to this distinction.
By avoiding the difficult manual orchestration required to scale PyTorch across GPUs, SPMD allows JAX to easily scale workloads across thousands of TPUs. JAX naturally offers a scaling advantage for large transformer models that need to be divided across multiple chips.
As a result, top research teams, such as those at Midjourney and Cohere, have moved their work to TPUs and JAX, citing significant cost savings and better performance indicators.
OpenAI Triton
Perhaps the most dangerous development for Nvidia is OpenAI’s Triton language.
Researchers can write extremely efficient GPU code with Triton without having to understand CUDA. It is an open-source language that abstracts the intricacies of hardware, much like Python.
Crucially, Triton is designed to be portable.
Red Hat is actively working to democratise AI accelerators with Triton, and it is already optimised for AMD and other accelerators. Triton successfully democratises the “secret sauce” that kept developers confined to Nvidia’s ecosystem by lowering the entry barrier for creating high-performance kernels. It transforms the GPU from a proprietary platform into a commodity compute unit.
The switching cost that safeguards Nvidia’s profits vanishes if a developer can write a kernel in Triton and have it operate efficiently on an AMD MI300, an Nvidia H100, or a future Google chip.
MLIR: The Infrastructure of Heterogeneity
The MLIR (Multi-Level Intermediate Representation) compiler infrastructure, which was first created at Google, is largely responsible for this change. MLIR lessens the fragmentation that previously afflicted non-Nvidia hardware by enabling various hardware vendors to create compilers that speak a common language.
The N×M problem, in which N frameworks (TensorFlow, PyTorch, JAX) must operate on M hardware backends (Nvidia, AMD, TPU, Trainium), is resolved by MLIR. The industry can create compilers that target a common MLIR representation in place of N×M compilers.
For rivals like Google and Amazon, this shared infrastructure significantly reduces the cost of developing reliable software stacks for their unique chips. In the future, AI workloads will be able to move smoothly across a heterogeneous hardware landscape thanks to this invisible plumbing.
The Smart Money Rotation
A quiet but significant rotation among sophisticated institutional managers is revealed by quantitative analysis of 13F filings, indicating a loss of faith in the “Nvidia only” trade.
The Hedge Fund Signal
The largest hedge fund in the world, Bridgewater Associates, reduced its Nvidia holdings by about 65% in Q3 2025 in order to reallocate funds to less volatile assets and broader indices. This action raises the possibility that, at current valuation levels, Nvidia’s risk/reward profile has substantially declined.
Stanley Druckenmiller, a renowned investor, sold his stake in Nvidia. He justified this by pointing to high valuations and the capital cycle’s predictable turn. His basic idea is still true even though he acknowledges that the sale was made too soon because the stock continued to rise. The era of effortless gains is over, and there will unavoidably be a period of digestion for the infrastructure expansion.
The move by Druckenmiller to Taiwan Semiconductor is instructive. In contrast to the design monopoly, which is currently facing increased competition, he is giving priority to the manufacturing monopoly that serves all of the major players, including Nvidia, Google, AMD, and Apple. This deliberate repositioning shifts funds away from the industry that controls access to the vital infrastructure.
A substantial $4.3 billion investment in Alphabet (Google) was started by Warren Buffett’s Berkshire Hathaway. This implies that the integrated incumbent with a strong moat (Search + Cloud + Silicon) is preferred over the high-flying “picks and shovels” vendor trading at peak multiples in a traditional value signal. Buffett’s investment in Google is both a subtle vote against the sustainability of the current chip mania and a confirmation of the company’s long-term resilience.
The Cisco Parallel
It is still foolish to dismiss the 2000 Cisco Systems comparison.
Important directional insight is provided by that parallel. As the backbone of the internet in 2000, Cisco was valued at more than 100 times its earnings. The business did not vanish. For years following the market correction, revenue kept rising. But when rivals appeared and the scarcity premium disappeared, the valuation multiple drastically shrank.
Nvidia’s present valuation presupposes perpetual exponential growth and margin resilience.
An impending multiple compression is strongly suggested by the emergence of powerful alternatives like Google’s TPU, AMD’s MI300, and Intel’s Gaudi, as well as the possible absorption of substantial capital expenditure.
Even if revenue growth continues, this compression is probably going to occur.
Despite market expansion, investors would suffer significant capital losses if Nvidia’s price-to-earnings ratio shrank from forty times to a more typical hardware valuation of twenty times.
The Four-Customer Problem
Nvidia’s revenue structure harbors a significant concentration risk.
This vulnerability is currently obscured by the rapid acceleration of its top-line expansion. While quarterly revenue has swelled to nearly $60 billion, the quality of these earnings remains fundamentally delicate. Data indicates that approximately 61 percent of this massive influx is derived from only four major counterparties. Microsoft, Amazon, Google, and Meta constitute a client base that is simultaneously complex and contradictory.
This concentrated reliance creates a precarious dynamic.
These four major customers are simultaneously Nvidia’s fiercest competitors.
Each is now developing proprietary silicon specifically to supplant Nvidia’s offerings.
These clients, like Google with Trillium, have the power to significantly restrict Nvidia’s future growth as soon as they have enough faith in their internal infrastructure. The company’s most significant sources of income are also the biggest danger to its survival.
Reports confirm that Gemini 3 was trained entirely on TPUs.
This is a crucial sign. Nvidia’s technology has been completely circumvented by the internal pipeline used to create one of the most advanced artificial intelligence models in the world.
At the cutting edge of the industry, the definitive proof of concept for “life without Nvidia” has now been established.
Geopolitics & Sovereign Clouds
For Nvidia, a company that sells tangible goods, U.S. export restrictions on cutting-edge AI chips to China and the Middle East present a special challenge. Nvidia was forced to create neutered versions like the H20 in order to comply with the Department of Commerce’s aggressive restrictions on the sale of high-performance chips to these areas.
Huawei’s Ascend series and other domestic Chinese alternatives are gaining market share, and even these compliant chips are now subject to restrictions.
The Export Control Asymmetry
Google’s operational paradigm is fundamentally distinct. The company monetises cloud access rather than physical hardware sales.
The process of selling a service, specifically API access to Gemini running on proprietary TPUs in a US data center, currently involves fewer transactional obstacles than the cross-border shipment of restricted silicon hardware, even as regulatory scrutiny of model weights and cloud compute intensifies.
As a result, Google can more easily profit from the demand for AI worldwide than Nvidia can from hardware sales in geographically restricted areas. Google can provide services to clients in the global south and other restricted jurisdictions thanks to the “Cloud Loophole” without having to transfer sensitive hardware.
This asymmetry becomes more significant to relative revenue growth as AI demand spreads throughout the world.
The Closed Garden Advantage
Nvidia is a fabless design house. It relies entirely on TSMC for manufacturing and a complex web of partners for packaging and assembly.
Google is also fabless, yet it designs its entire stack. This includes the datacenter rack, cooling systems, and networking fabric, known as Jupiter. This complete vertical integration allows Google to optimise for supply chain resilience.
Importantly, Optical Circuit Switching is a proprietary technology used in Google’s TPU Pods. This decision lessens their reliance on costly and power-hungry InfiniBand switches.
High profits are extracted from this same sector by Nvidia. Google shields itself from the supply shortages and exorbitant prices that afflict the merchant networking industry by having complete control over the interconnect layer.
The environment of Gemini is like a “closed garden”.
Google strategically extracts value from all layers, including the proprietary model, cloud, software, and silicon.
Google’s R&D investment will yield faster returns thanks to this integration, which guarantees that each component improves the others. Importantly, this integrated strategy concurrently reduces revenue prospects for outside suppliers like Nvidia.
The Post-GPU Equilibrium
The investment cycle in artificial intelligence continues, but the oversimplified strategy of only supporting Nvidia is now out of date.
A sophisticated era characterised by a variety of architectures and fierce competition focused on unit economics is upon us.
The deployment of capital is undergoing a fundamental shift.
The Probability Matrix
Our internal scenario analysis at Bancara assigns the following probabilities to potential market structures by 2027:
Bull Case (20% probability): Artificial intelligence is still in high demand. The dominance of CUDA seems to be enduring. The market penetration of custom silicon initiatives is still a challenge. As a result, Nvidia continues its hyper-growth trajectory and maintains remarkable profit margins.
Base Case (50% probability): The market is breaking apart. Thirty to forty percent of hyperscalers’ workloads will be successfully moved to proprietary silicon by 2026. Nvidia may lose the high-volume inference market even though it will keep the premium training segment. As a result, it is anticipated that operating margins will narrow to between 55 and 60 percent.
Bear Case (30% probability): Different hardware becomes functionally equivalent thanks to software abstraction using technologies like Triton and JAX. The competitive environment created by Trainium and TPU will drastically lower inference costs. As a result, Nvidia will only be one of many suppliers of components in the chip industry.
The Mechanism of Decline
Why exactly is Google’s win bad for Nvidia stock?
The mechanism operates through three channels:
Direct Revenue Hit: A lost H100 sale is represented by each deployed TPU. Billions of dollars in potential revenue from Nvidia’s forecast are being directly eroded by Google’s deployment of massive Trillium clusters.
Pricing Power Erosion: AWS and Azure are forced to lower their own prices in order to stay competitive due to Google’s aggressively priced inference. AWS and Azure must therefore either demand lower chip costs from Nvidia or accelerate the development and implementation of their own custom silicon solutions in order to protect their profit margins.
Multiple Contraction: The expectation of infinite growth fades as the market recognises limitations on Nvidia’s total addressable market as a result of hyperscaler vertical integration. For present investors, a valuation change from forty times sales to twenty times sales would be disastrous.
The Strategic Verdict
The TPU project at Google is more than just a side project.
It represents an essential safeguard against the democratisation of artificial intelligence.
By integrating control over the silicon, the software, the cloud and the foundational model Google has erected a formidable structure that proves impenetrable to Nvidia. The market is currently going through a major split. This division separates the low-cost domain-specific hyperscaler sector from the high-cost general-purpose merchant sector.
The laws of economics are repealing the “Nvidia Tax”. The vertically integrated giants who can generate intelligence at the lowest marginal cost will be the alpha in the next stage of the AI trade, not merchant arms dealers.
Bancara is aligning portfolios for the Second Phase of the AI revolution. This strategy involves rotating exposure away from margin-vulnerable merchant silicon towards vertically integrated platforms with highly defensible cost structures.
The investment thesis is unambiguous.
The Nvidia Tax is being repealed.
Position accordingly.
For information only; not investment advice or a solicitation.
Bancara Insights — Global perspective. Multi-asset access. Discreet service.
