NVIDIA's Unassailable Position: A Technical Analysis of Why the Moat Holds Through 2030

December 2025 Update: NVIDIA maintaining 80% AI accelerator share, 78% gross margins despite challengers (DeepSeek, TPUs, MI300X, export controls). Stock drops on threats, then recovers. Moat isn't CUDA itself—it's 19 years of accumulated ecosystem: cuDNN, cuBLAS, NCCL, PyTorch/TensorFlow optimization, Nsight toolchain, documentation. Switching costs exceed performance advantages for virtually every customer.

Every threat to NVIDIA follows the same script. Analysts identify a challenger—DeepSeek's efficiency, Google's TPUs, AMD's MI300X, open-source models, export controls—and predict market share erosion. The stock drops. Headlines multiply. Then the threat passes. Market share remains at 80%.¹ Gross margins hold at 78%.² The hyperscalers announce another round of capital expenditure, most flowing to NVIDIA hardware.³

The pattern repeats because the analysis focuses on the wrong variable. Observers compare specifications and conclude that competitors have caught up or will soon. The comparison misses what makes NVIDIA's position durable: switching costs that exceed performance advantages by such a margin that rational actors stay even when alternatives offer better specs.

NVIDIA will maintain dominant market share through 2030. Not because competitors won't produce better hardware on specific metrics—they already have in some cases. Not because efficiency gains won't reduce per-model compute requirements—they already have. NVIDIA wins because the total cost of switching platforms exceeds the total benefit of the switch for virtually every customer in the market. Understanding why requires understanding what the moat actually comprises.

The moat isn't CUDA. The moat is everything built on CUDA.

CUDA launched in 2006. Nineteen years of accumulated investment followed. That investment didn't just create a programming interface. It created an ecosystem so comprehensive that CUDA functions less like a software platform and more like the foundational infrastructure of AI development itself.

The base layer comprises the parallel computing model and programming abstractions. CUDA provides a way for developers to express parallel computations that execute efficiently on GPU architectures. This base layer works well, but it could theoretically be replicated. AMD's ROCm provides similar abstractions. Intel's oneAPI attempts the same.

The accumulated layers above the base create the defensible advantage.

Libraries and primitives: cuDNN for deep learning primitives. cuBLAS for linear algebra. cuFFT for Fourier transforms. Thrust for parallel algorithms. NCCL for multi-GPU communication. Each library represents thousands of engineering hours optimizing for NVIDIA architectures. Each optimization compounds with others. A model that uses cuDNN for convolutions, cuBLAS for matrix operations, and NCCL for gradient aggregation captures optimizations at every layer of the stack.⁴

Framework integration: PyTorch, TensorFlow, JAX, and every other major framework optimize first and most deeply for NVIDIA GPUs. The framework developers use NVIDIA hardware. The framework test suites run on NVIDIA hardware. Bug reports come primarily from NVIDIA users. The frameworks work on other hardware; they work best on NVIDIA hardware.⁵

Toolchains and debugging: Nsight for profiling and debugging. CUDA-GDB for kernel debugging. Compute Sanitizer for error detection. Tools that help developers write correct, efficient code. Tools that don't exist or exist in immature form for competing platforms.

Documentation and knowledge: Nineteen years of blog posts, tutorials, academic papers, Stack Overflow answers, and institutional knowledge. When a developer encounters a CUDA problem, the solution exists somewhere. When a developer encounters a ROCm problem, they might be the first person to see it.

Developer muscle memory: Graduate students learn CUDA. Research teams use CUDA. Engineers build careers around CUDA expertise. The people who make technology decisions have spent years accumulating CUDA-specific skills that don't transfer to other platforms.

The layers compound. An organization switching from NVIDIA to AMD doesn't just change hardware. It rewrites CUDA kernels to HIP or ROCm. It replaces cuDNN calls with MIOpen calls. It retrains developers. It abandons Nsight and learns new tools. It leaves behind the community knowledge that solves esoteric problems at 2 AM. It takes on debugging risk in an ecosystem with less coverage.

Each layer adds switching cost. The switching costs stack multiplicatively, not additively. A 20% advantage on paper becomes a 20% disadvantage in practice when achieving it requires rebuilding the entire stack from scratch.

Why DeepSeek proved the moat rather than threatening it

DeepSeek's January 2025 announcement claimed frontier AI models could be trained for $6 million instead of $600 million.⁶ The market interpreted this as an existential threat: if models could be built cheaply, demand for expensive hardware would collapse.

The interpretation failed on multiple levels, each revealing aspects of NVIDIA's structural strength.

Efficiency gains don't reduce demand; they expand it. Jevons Paradox—the observation that efficiency improvements increase rather than decrease total resource consumption—applies directly. When training costs drop by 99%, the addressable market expands by more than 99x. Organizations that couldn't afford frontier AI at $600 million can afford it at $6 million. The aggregate compute consumption increases even as per-model consumption decreases.

Meta's response demonstrated this immediately. Days after DeepSeek's announcement, Meta raised its 2025 AI spending guidance to $60-65 billion.⁷ The company saw cheaper training as a reason to train more models for more use cases, not a reason to reduce infrastructure investment.

DeepSeek ran on NVIDIA hardware. The company used export-restricted NVIDIA chips supplemented by Huawei's Ascend 910B, which achieves 91% of comparable NVIDIA performance.⁸ Even the company supposedly threatening NVIDIA's dominance couldn't escape NVIDIA's ecosystem entirely. The efficiency innovations DeepSeek developed—mixture of experts, attention optimization, training curriculum improvements—transfer to NVIDIA hardware. Organizations that want DeepSeek's efficiency can achieve it while staying on NVIDIA's platform.

The market correctly processed the signal within 48 hours. NVIDIA's $593 billion single-day loss reversed as institutional investors recognized the overreaction.⁹ The stock recovered 8.9% the following day. Retail investors sold; institutions bought the dip. The sophisticated market participants understood what the headlines missed.

Industrial commitment didn't waver. Chevron and GE Vernova announced plans to build dedicated power plants for data centers after DeepSeek's announcement, not before.¹⁰ Industrial companies don't commit billions to infrastructure projects based on bubbles or soon-to-be-obsolete technologies. They build for decades of sustained demand.

The DeepSeek episode tested NVIDIA's moat with the most favorable possible conditions for the bear case: dramatic efficiency improvements, from a competitor unconstrained by US export regulations, announced at peak market exuberance. The moat held. Any future challenge operates under less favorable conditions.

TPUs: real competition in a defined segment, not a platform threat

Google's Tensor Processing Units represent genuine competition. TPUv7 (Ironwood) delivers 4,614 TFLOPS in BF16, a 10x improvement over TPUv5p.¹¹ Google has won significant customers: Anthropic's buildout exceeds 1 GW of TPU capacity.¹² Meta reportedly plans to use TPUs in data centers by 2027.¹³ OpenAI, SSI, and xAI have discussed TPU access with Google.¹⁴

The wins are real. They don't threaten NVIDIA's dominant position because they occur in a specific market segment with characteristics that don't generalize.

TPUs optimize for inference cost at hyperscale. Inference costs for production AI systems exceed training costs by 15-118x.¹⁵ At hyperscale, inference cost optimization drives significant economic value. Google's TPUs deliver 4.7x better performance per dollar and 67% lower power consumption for these workloads.¹⁶ For organizations running inference at massive scale with cost as the primary constraint, TPUs offer compelling economics.

TPUs remain captive to Google's ecosystem. Organizations access TPUs through Google Cloud or through direct relationships with Google. The hardware doesn't ship to customer data centers. The software ecosystem doesn't exist independently of Google's infrastructure. Choosing TPUs means choosing Google as a strategic partner at a fundamental level.

This constraint eliminates most of the market. Enterprises deploying AI in their own data centers cannot use TPUs. Organizations unwilling to concentrate infrastructure with a single hyperscaler cannot use TPUs. Companies in regulated industries that prohibit specific cloud dependencies cannot use TPUs. The constraint doesn't apply to Anthropic or Meta, which operate at sufficient scale to negotiate direct relationships. It applies to the long tail of the market.

Training still happens predominantly on NVIDIA. Google trains Gemini on TPUs. Everyone else trains on NVIDIA. The training market differs from the inference market in several ways: training workloads are more varied and less standardized than inference; training requires more flexibility to experiment with architectures; training benefits more from ecosystem depth. NVIDIA's position in training remains stronger than its position in inference.

Market segmentation doesn't equal market loss. If TPUs capture 20% of hyperscale inference while NVIDIA retains 95% of training, 90% of enterprise inference, and 80% of other hyperscale inference, NVIDIA's absolute volume and revenue continue growing. The AI compute market expands faster than any segment TPUs might capture. NVIDIA's share could decline slightly while its revenue doubles.

The prediction: TPUs become a meaningful part of the AI compute landscape, specifically for cost-sensitive inference at hyperscale. NVIDIA retains training dominance, enterprise dominance, and a majority of hyperscale compute. Both companies grow. The framing of TPUs as an NVIDIA "threat" mistakes segment competition for platform displacement.

AMD MI300X: specifications win benchmarks, ecosystems win markets

AMD's MI300X offers compelling specifications: 192 GB of HBM3 memory versus 80 GB for the H100.¹⁷ For memory-bound inference workloads, more memory matters. Large language models during inference often bottleneck on memory bandwidth rather than compute. The MI300X specification sheet presents genuine competitive hardware.

Market share tells a different story. Omdia estimates NVIDIA holds approximately 80% of the AI accelerator market.¹⁸ AMD captures single-digit percentage points. The gap has not closed meaningfully despite multiple generations of competitive hardware releases.

The pattern extends across AMD's entire competitive history with NVIDIA. Each generation, AMD announces hardware that matches or exceeds NVIDIA on specifications. Each generation, NVIDIA maintains market share. Each generation, observers predict the gap will close. Each generation, it doesn't.

The consistency of this pattern across fifteen years of competition provides strong evidence that something other than specifications determines market outcomes. That something is the ecosystem.

ROCm, AMD's answer to CUDA, exists and functions. Framework support exists. Libraries exist. Documentation exists. But each element exists at lower density than the NVIDIA equivalent. PyTorch works on ROCm; more PyTorch users run on CUDA. MIOpen provides deep learning primitives; cuDNN provides more optimized primitives for more architectures. Community knowledge accumulates more slowly because fewer developers encounter and solve ROCm problems.

ROCm launched in 2016—nine years ago. AMD has invested consistently in the platform. The technical foundation works. The gap persists because ecosystems exhibit network effects. Developers optimize for CUDA because users run CUDA. Users run CUDA because developers optimize for it. Breaking the cycle requires simultaneous movement by both groups that neither has incentive to initiate.

The investment disparity compounds. AMD's R&D budget runs roughly $5.8 billion annually across all products—CPUs, GPUs, embedded systems, and software.²⁷ NVIDIA spent $8.7 billion on research and development in fiscal 2024, increasingly focused on AI software and ecosystem.²⁸ The spending gap represents years of accumulated advantage in libraries, tools, documentation, and developer relations.

More fundamentally: NVIDIA treats CUDA as its primary competitive advantage and invests accordingly. AMD treats ROCm as necessary infrastructure to sell GPUs—important but not primary. The strategic priority difference shows in execution quality, documentation depth, developer relations investment, and release cadence. Closing the ecosystem gap would require AMD to sustain above-parity investment in software for 5+ years while accepting below-parity returns until the ecosystem reaches critical mass. The business case doesn't support this strategy.

AMD's rational approach is to compete for specific segments where ecosystem matters less—high-memory inference, traditional HPC workloads, customers with existing AMD relationships—rather than attempting ecosystem displacement. This strategy can work for market share in single digits. It cannot close the gap with NVIDIA.

The practical consequence: developers optimize for CUDA first, then port to ROCm if business requirements demand it. The porting process takes time, introduces bugs, and often results in performance degradation. Organizations evaluating both platforms find that the NVIDIA path involves less risk, faster deployment, and better support.

AMD faces additional pressure from TPUs. When Meta reportedly considered Google's TPUs, AMD shares fell 6%.¹⁹ AMD occupies an uncomfortable competitive position: similar enough to NVIDIA to seem like a substitute but not distinct enough to capture a unique segment. TPUs offer something architecturally different optimized for a specific use case. MI300X offers something similar to NVIDIA that requires ecosystem sacrifice to adopt.

The MI300X and its successors will win specific deployments where memory capacity constraints dominate other considerations. AMD will not close the market share gap with NVIDIA through hardware specifications alone because the gap isn't about hardware specifications.

The full-stack compounding effect

NVIDIA's advantage extends beyond CUDA into a full-stack integration that creates multiple layers of lock-in and switching cost.

Hardware innovations compound with software. Blackwell architecture introduced FP4 precision for inference, enabling 4x compute density for workloads that can use lower precision.²⁰ The FP4 capability requires hardware support. It also requires software support: libraries that know how to use FP4, frameworks that expose it properly, applications that can adapt to lower precision. NVIDIA ships both the hardware capability and the software stack to use it. Competitors must replicate both.

Networking creates additional lock-in. NVLink provides 900 GB/s GPU-to-GPU bandwidth, 7x faster than PCIe.²¹ NVSwitch enables all-to-all GPU communication within a node. InfiniBand (via NVIDIA's Mellanox acquisition) provides inter-node communication optimized for GPU workloads. The networking stack integrates with the computing stack at every level. Moving GPUs often means moving networking infrastructure.

Systems integration captures additional value. DGX systems provide turnkey infrastructure: servers, networking, storage, and software configured to work together. Organizations that buy DGX systems lock in at the system level, not just the GPU level. The integration saves deployment time while adding switching cost.

Software platforms extend the lock-in. NVIDIA AI Enterprise provides deployment tools, optimization frameworks, and management capabilities. The platform creates switching costs independent of the hardware. An organization using AI Enterprise has workflows, scripts, and operational procedures built around NVIDIA's platform software.

Startup ecosystem compounds over time. NVIDIA's 2025 investments included 49 AI startups; the Inception Program includes over 15,000 participants.²² Startups build on CUDA because NVIDIA makes building on CUDA easy. By the time those startups scale, switching costs have accumulated across their entire technology stack. The ecosystem investment creates future customers.

Each layer reinforces the others. Moving GPUs requires evaluating networking impact. Moving networking affects systems integration. Moving systems integration disrupts software platforms. Moving software platforms invalidates startup ecosystem benefits. The switching cost isn't the sum of individual layer costs; it's the cost of coordinating changes across all layers simultaneously.

What switching actually costs

Abstract discussion of switching costs obscures the practical reality. Consider what an organization actually faces when evaluating a move from NVIDIA to AMD.

The hypothetical: A mid-sized AI company runs inference on 500 NVIDIA A100 GPUs. AMD offers MI300X with 192 GB HBM3 versus A100's 80 GB HBM2e—more than double the memory at comparable price. The specification advantage is real. Memory-bound LLM inference would benefit measurably.

Month 1-2: Evaluation. The team acquires test hardware. Engineers discover ROCm version incompatibilities with existing PyTorch code. They identify custom CUDA kernels that require manual porting. Initial estimate: 15-20% of the codebase needs modification. The percentage seems manageable.

Month 3-4: Porting. Engineers rewrite CUDA kernels to HIP, AMD's CUDA translation layer. They encounter edge cases where ROCm behavior differs from CUDA. Memory management issues appear that didn't exist on NVIDIA hardware. The ported code runs 10-20% slower than the CUDA equivalent, despite the MI300X's theoretical specification advantage.

Month 5-6: Optimization. The team profiles the ported code to identify bottlenecks. They discover ROCm profiling tools are less mature than NVIDIA's Nsight ecosystem. Some optimizations that worked on CUDA don't apply to AMD's architecture. New architecture-specific optimizations require expertise the team doesn't have. They consider hiring AMD-specialized engineers and find the talent pool is shallow.

Month 7-8: Production qualification. The team runs A/B tests against the existing NVIDIA fleet. Production edge cases appear that testing didn't cover. Debugging these issues proves difficult—Stack Overflow has 50 times more CUDA questions than ROCm questions.²⁶ The community knowledge that solves obscure problems at 2 AM doesn't exist for AMD.

Month 9 and beyond: Maintenance. The organization now maintains two codepaths during the transition. When NVIDIA releases new features—FP8 precision, TensorRT-LLM optimizations—engineers must decide whether to implement on one platform or both. Team cognitive load increases. New hires require training on two platforms instead of one.

The opportunity cost: During 6-9 months of migration work, competitors shipping on NVIDIA continue improving their products. The switching organization spends engineering resources on platform migration rather than product development. Even if the migration succeeds completely, the opportunity cost may exceed the operational savings from better hardware specifications.

The calculation that keeps organizations on NVIDIA: Most organizations conclude that a 20-30% operational cost advantage from AMD doesn't justify 6-12 months of migration work, 10-20% performance regression during transition, ongoing maintenance burden, talent acquisition challenges, and competitive opportunity cost. They stay on NVIDIA and allocate engineering resources to product development instead.

This calculation repeats across thousands of organizations. Each concludes independently that switching doesn't make sense. The aggregate effect: AMD's market share remains in single digits despite competitive hardware.

Why export controls don't change the analysis

US export restrictions have cut NVIDIA off from the Chinese market. The H20, designed to comply with earlier restrictions, faced additional licensing requirements in April 2025.²³ NVIDIA reported $2.5 billion in missed H20 revenue during Q1 2025 and expected another $8 billion loss in Q2.²⁴ Total regulatory impact could reach $15-16 billion annually.

The numbers matter. China represented 15-20% of NVIDIA's revenue before restrictions. Losing that revenue hurts the company.

The numbers don't change the competitive analysis for the non-China market.

Domestic and allied markets contain the vast majority of AI investment. The $380 billion in hyperscaler capital expenditure flows to facilities in the United States and allied nations.²⁵ European AI investment continues. Asian allies—Japan, South Korea, Taiwan—remain accessible markets. The markets where NVIDIA competes with AMD, TPUs, and other alternatives are the markets where NVIDIA retains full access.

Export controls may strengthen NVIDIA's long-term competitive position outside China. Chinese organizations must now develop domestic alternatives without access to leading-edge NVIDIA technology. Those alternatives—Huawei's Ascend chips, other domestic developments—will lag NVIDIA's trajectory because they cannot access NVIDIA's accumulated innovations. When restrictions eventually ease, Chinese organizations may face a choice between inferior domestic hardware and superior NVIDIA hardware with years of additional development.

The China market is lost for now. The rest of the world—representing 80-85% of AI infrastructure investment—remains NVIDIA's to win or lose based on competitive dynamics. Nothing in those dynamics suggests meaningful share loss.

What would actually change this

Intellectual honesty requires identifying what could erode NVIDIA's position, even if those conditions seem unlikely near-term.

Scenario 1: True hardware abstraction. If PyTorch or another framework achieved genuine hardware abstraction—where code runs identically on NVIDIA, AMD, TPU, and other accelerators without modification—the switching cost equation changes fundamentally. Organizations could migrate gradually without rewriting code. Developers could optimize once and deploy anywhere.

This hasn't happened despite years of effort. Hardware differences in memory models, synchronization primitives, and performance characteristics leak through abstractions. XLA attempts this for TPUs with partial success. OpenAI's Triton provides a higher-level abstraction that targets multiple backends. The technical challenge is substantial: hardware architectures differ in ways that matter for performance, and abstracting those differences away often means sacrificing performance. Progress is slow. True hardware abstraction remains years away if achievable at all.

Scenario 2: Architectural discontinuity. A fundamentally new computing paradigm—optical computing, neuromorphic chips, or practical quantum systems for relevant AI workloads—could obsolete the GPU computing model entirely. In this scenario, NVIDIA's accumulated advantages become liabilities: expertise in an obsolete architecture, ecosystem lock-in to deprecated abstractions, organizational inertia resisting the new paradigm.

Timeline: 10+ years minimum for any of these to achieve AI-relevant scale. Optical computing shows promise for inference but remains laboratory-bound. Neuromorphic chips serve niche applications. Quantum computing addresses different problem classes than current AI workloads. NVIDIA actively researches all of these technologies.²⁹ The company would likely adapt to an architectural shift rather than be displaced by it—just as NVIDIA transitioned from graphics to general-purpose computing to AI.

Scenario 3: Sustained alternative ecosystem investment. A major player—Google, Microsoft, Meta, or a nation-state—could fund a sustained 5-10 year effort to build a CUDA-competitive ecosystem around an alternative architecture. Google's approach with TPUs and JAX represents a partial version of this strategy. China's domestic semiconductor effort attempts something similar under export control pressure.

The challenge: the effort would require billions in annual investment with returns deferred for years. It would require coordinating framework developers, library maintainers, hardware vendors, and enterprise adopters simultaneously. It would require patience that quarterly earnings pressure discourages. Google comes closest to having both the resources and the strategic motivation, but even Google's TPU ecosystem remains captive to Google Cloud rather than open competition with CUDA.

What remains stable: None of these scenarios operates on the 2025-2030 timeframe. Hardware abstraction advances incrementally. Architectural discontinuities require breakthrough discoveries that haven't occurred. Sustained ecosystem investment requires strategic patience that market structures discourage. The conditions that would threaten NVIDIA's position exist in theory. They don't exist in practice on any timeline relevant to infrastructure decisions made today.

The 2030 view

NVIDIA will maintain dominant market share through 2030 because:

1. Switching costs exceed performance advantages. Even if AMD or another competitor produces hardware 30% faster on relevant benchmarks, the total cost of switching—rewriting code, retraining developers, abandoning ecosystem advantages, accepting support and debugging risk—exceeds the performance benefit for virtually every organization in the market.

2. The full-stack compounds. Each layer of NVIDIA's stack creates independent switching costs that multiply together. Displacing NVIDIA requires winning at every layer simultaneously. No competitor is positioned to do this.

3. TPUs compete in a segment, not for the platform. Google will capture hyperscale inference share. NVIDIA will retain training dominance, enterprise dominance, and the majority of compute across all segments. Both absolute markets grow.

4. AMD cannot close the ecosystem gap through hardware. Fifteen years of history demonstrate that better specifications don't translate to market share gains when the ecosystem advantage favors the incumbent.

5. Efficiency gains expand the market. Every efficiency improvement that reduces per-model compute requirements expands the addressable market by making AI viable for more organizations. NVIDIA benefits from market expansion regardless of the specific technique that enables it.

The competitive landscape in 2035 may differ. Architectural shifts, sustained alternative ecosystem development, or regulatory changes could eventually erode NVIDIA's position. But "eventually" operates on a longer timeline than most analysis assumes.

Betting against NVIDIA dominance through 2030 requires believing that switching costs suddenly decrease, that ecosystem dynamics suddenly reverse, or that the full-stack advantage suddenly stops compounding. Nothing in the competitive environment suggests any of these conditions will emerge.

The moat holds.

Key takeaways

For strategic planners: - NVIDIA maintains 80% market share through 2030; 19 years of CUDA ecosystem investment creates switching costs exceeding any competitor's performance advantage - $380B hyperscaler CapEx flows predominantly to NVIDIA; Meta raised 2025 AI spending to $60-65B post-DeepSeek announcement - TPUs compete in hyperscale inference segment (Anthropic 1GW+), not for platform displacement; training remains NVIDIA-dominated

For finance teams: - NVIDIA gross margins hold at 78%; export control impact: $2.5B Q1 2025 missed H20 revenue, $8B expected Q2 loss, $15-16B annually - DeepSeek's $6M training claim triggered $593B single-day loss—stock recovered 8.9% next day as institutions recognized overreaction - Efficiency gains expand markets (Jevons Paradox): cheaper training enables more organizations to afford AI, increasing aggregate compute demand

For engineering leadership: - Platform migration requires 6-12 months: CUDA kernel rewrites, cuDNN→MIOpen replacements, developer retraining, Nsight abandonment - 10-20% performance regression during transition despite AMD's specification advantages; Stack Overflow has 50x more CUDA than ROCm questions - ROCm launched 2016—nine years of investment hasn't closed ecosystem gap; AMD R&D $5.8B vs NVIDIA $8.7B focused on AI software

For competitive intelligence: - AMD MI300X offers 192GB HBM3 vs H100's 80GB; 15 years of competitive hardware releases never translated to market share gains - TPUv7 delivers 4,614 TFLOPS BF16 (10x TPUv5p); 4.7x better $/performance for inference but remains captive to Google Cloud - True hardware abstraction or architectural discontinuity (optical, neuromorphic, quantum) could threaten position but operates on 10+ year timeline

Infrastructure deployments require expertise across GPU architectures, networking, and systems integration. Teams at Introl deploy AI infrastructure across 257 global locations, helping organizations navigate hardware decisions with the technical depth that platform choices require.

References

1. Seeking Alpha. "Google Vs. Nvidia: Trillions of Dollars at Play." Seeking Alpha, 2025.

2. CNBC. "Nvidia Sales Are 'Off the Charts,' but Google, Amazon and Others Now Make Their Own Custom AI Chips." CNBC, November 21, 2025.

3. ———. "How Much Google, Meta, Amazon and Microsoft Are Spending on AI." CNBC, October 31, 2025.

4. NVIDIA Developer. "CUDA Libraries Documentation." NVIDIA Developer Zone, 2025.

5. PyTorch. "PyTorch Performance Tuning Guide." PyTorch Documentation, 2025.

6. Fortune. "What Is Jevons Paradox? The Reason Satya Nadella Says DeepSeek's New AI Is Good News for Tech." Fortune, January 27, 2025.

7. Substack. "The Jevons Paradox in AI Infrastructure: DeepSeek Efficiency Breakthroughs to Drive Energy Demand." AI Proem, 2025.

8. Agility PR Solutions. "The DeepSeek Disruption: NVIDIA's Biggest AI Challenge Yet?" Agility PR Solutions, 2025.

9. The Motley Fool. "Does DeepSeek's Breakthrough Help or Hurt Nvidia Stock? 3 Views You Should Know." The Motley Fool, February 3, 2025.

10. Axios. "DeepSeek Illustrates AI's Jevons Paradox." Axios, January 29, 2025.

11. SemiAnalysis. "Google TPUv7: The 900lb Gorilla in the Room." SemiAnalysis Newsletter, 2025.

12. ———. "Google TPUv7."

13. CNBC. "Nvidia Stock Falls 4% on Report Meta Will Use Google AI Chips." CNBC, November 25, 2025.

14. SemiAnalysis. "Google TPUv7."

15. AI News Hub. "Nvidia to Google TPU Migration 2025: The $6.32B Inference Cost Crisis." AI News Hub, 2025.

16. ———. "Nvidia to Google TPU Migration 2025."

17. AMD. "AMD Instinct MI300X Specifications." AMD Corporation, 2024.

18. Seeking Alpha. "Google Vs. Nvidia."

19. CNBC. "Nvidia Stock Falls 4%."

20. NVIDIA. "Blackwell Architecture Technical Brief." NVIDIA Corporation, 2024.

21. NVIDIA. "NVLink and NVSwitch Technical Overview." NVIDIA Data Center Solutions, 2025.

22. Klover.ai. "NVIDIA AI Strategy: Analysis of Sustained Dominance in AI." Klover.ai, 2025.

23. FXOpen. "Analytical Nvidia (NVDA) Stock Projections for 2025-2040." FXOpen Blog, 2025.

24. ———. "Analytical Nvidia (NVDA) Stock Projections."

25. CNBC. "How Much Google, Meta, Amazon and Microsoft Are Spending on AI."

26. Stack Overflow. "Tag Statistics: CUDA vs ROCm." Stack Overflow, 2025.

27. AMD. "AMD Annual Report 2024: Research and Development." AMD Investor Relations, 2024.

28. NVIDIA. "NVIDIA Fiscal 2024 Annual Report." NVIDIA Investor Relations, 2024.

29. NVIDIA. "NVIDIA Research: Exploring Next-Generation Computing Paradigms." NVIDIA Research, 2025.