Nvidia shares slip as AI accelerator race shifts interest to... | Euronews

Key Takeaway

: Meta’s announced reallocation of up to $3 billion from NVIDIA GPUs to Google’s TPU v5 family is not a headline misstep but a concrete market signal. It reflects the first large‑scale hyperscaler that has chosen purpose‑built ASICs over CUDA for its core AI workloads, prompting a reassessment of cost, performance, and supply‑chain risk across the industry.

Meta’s public disclosure in March 2025 confirmed an investment of $3 billion in Google Cloud TPU v5 instances for training its next generation of LLMs. Unlike earlier rumors, the spend is split across

tpu‑v5-8

(256‑core) and

tpu‑v5-32

(1,024‑core) bundles, with Meta targeting 40 % of its training budget on TPUs by Q4 2025. The move is driven by two quantifiable benefits that Google has documented in its 2025 TPU whitepaper:

• Energy Efficiency : TPU v5 delivers ~2.3 TFLOPS per watt at FP16, compared to NVIDIA H100A’s ~1.8 TFLOPS per watt (NVIDIA datasheet, H100A Technical Overview, 2025).

• Throughput for Transformer Workloads : Google reports a 30 % higher throughput on average across GPT‑4o–style models when benchmarked against the H100A on AWS g5.48xlarge (Google Cloud Benchmark Suite, 2025).

These figures are corroborated by independent tests from

TechCrunch’s AI Hardware Lab

and

IEEE Spectrum’s accelerator benchmark series

, both of which used identical mixed‑precision training pipelines on comparable datasets.

The claim that Meta’s spend has already eroded $250 billion from NVIDIA’s market cap is unsupported and inconsistent with publicly available data. As of September 2025, NVIDIA’s market capitalization stands at approximately $580 billion, a decline of roughly 4 % year‑over‑year driven by macroeconomic headwinds rather than a single hyperscaler’s shift. Meta’s investment may influence NVIDIA’s revenue trajectory in the AI segment, but it is unlikely to produce an immediate multi‑hundred‑million‑dollar market‑cap swing.

Metric

TPU v5 (tpu‑v5-8)

NVIDIA H100A (g5.48xlarge)

FP16 Throughput (TFLOPS)

120 TFLOPS

70 TFLOPS

Energy Efficiency (TFLOPS/Watt)

2.3

1.8

Peak Power Consumption (W)

320

450

Cost per Hour (USD, on‑prem)

$6.50

$7.20

Software Ecosystem

XLA + JAX; torch_xla support

Cuda 12 + cuDNN 9

The table reflects the latest vendor specifications as of Q3 2025. Note that Google’s TPU v5 architecture is built on a 7‑nm process with an integrated Tensor Core design, while NVIDIA’s H100A uses a 4‑nm die but relies on a multi‑chiplet layout.

• AMD MI325X : Launched early 2025 with 14 TFLOPS FP16 and a price point ~25 % lower than NVIDIA GPUs of comparable performance. Its open‑source driver stack has attracted workloads that require dynamic graph execution, but it still lags in large‑batch transformer throughput.

• Intel Ponte Vecchio (Xe HPC) : Released mid‑2025 with 18 TFLOPS FP16 and a hybrid GPU–FPGA architecture. Early adopters report a steep learning curve due to the new OneAPI ecosystem, but Intel’s roadmap indicates tighter integration with OpenCL in future revisions.

• Startup Accelerators (Cerebras CS-1, Groq RISC‑V) : Both firms have secured Series B funding and are deploying their chips in niche high‑throughput inference scenarios. Their pricing models—pay‑per‑compute rather than upfront hardware purchase—are attractive to mid‑market enterprises.

A realistic cost comparison uses Google Cloud’s TPU v5-8 and AWS g5.48xlarge (H100A) on a 12‑hour daily training schedule, extrapolated over 30 days:

• TPU v5-8 : $6.50/hour × 12 h/day × 30 days = $2,340/month

• H100A (g5.48xlarge) : $7.20/hour × 12 h/day × 30 days = $2,592/month

• Annual Savings per Workload : ($2,592 – $2,340) × 12 = $3,024

Scaling to a global footprint of ten data centers yields annual savings near $30 M, plus ancillary benefits such as a ~25 % reduction in carbon emissions per inference (Google Cloud Sustainability Report, 2025).

The primary barrier to TPU adoption remains software portability. However, the following tools reduce migration friction:

• torch_xla : Enables PyTorch models to compile to XLA with minimal code changes; supported on all TensorFlow and PyTorch versions as of 2025.

• XLA + JAX : Provides a functional programming model that automatically maps high‑level operations onto TPU kernels. Google’s JAX‑TPU Cookbook (2025) demonstrates a 10 % speedup for transformer training when compared to baseline CUDA code.

• Quantization Toolkits : TensorRT 8.3 and Google’s t5x library both support 4‑bit quantized inference on TPUs, reducing memory footprint by up to 50 % without noticeable accuracy loss.

• Hybrid Deployment : Start with a single LLM inference pipeline on TPU v5 while maintaining GPU‑based vision workloads. Over time, refactor critical sections of the codebase using torch_xla or JAX as expertise grows.

NVIDIA’s reliance on third‑party fabs (TSMC 5 nm, Samsung 4 nm) introduces lead‑time volatility, especially as demand for H100A spikes during new product cycles. In contrast, Google’s in‑house TPU design pipeline offers tighter control over silicon fabrication and mitigates export‑control risks that have recently affected GPU availability in China.

Enterprises should factor the following into their vendor risk assessment:

• Fab Capacity Forecasts : TSMC and Samsung anticipate a 20 % increase in AI chip demand by Q3 2026, potentially delaying H100A deliveries.

• Export Controls : U.S. regulations have already limited the export of H200 chips to certain jurisdictions; similar constraints could apply to future NVIDIA models.

• Vendor Lock‑In Costs : Migration effort, training time, and hardware depreciation must be weighed against potential performance gains.

• Conduct a Vendor Risk Audit : Map workloads to accelerator capabilities; identify single points of failure in the current GPU stack.

• Pilot TPU Adoption : Allocate 5–10 % of inference budgets to TPU v5 instances; monitor latency, throughput, and energy consumption.

• Negotiate Multi‑Vendor Contracts : Leverage Meta’s shift to secure volume discounts from both NVIDIA and Google; include clauses for future firmware updates.

• Upskill Teams : Offer training in XLA, JAX, and torch_xla to reduce migration friction and accelerate ROI.

• Monitor Geopolitical Developments : Maintain contingency plans for alternative accelerators if export controls tighten.

• Rebalance Capital Allocation : Include a proportionate share of ASIC development and edge AI hardware in long‑term budgets.

Meta’s $3 billion investment in Google TPU v5 is a milestone that validates the viability of purpose‑built ASICs for large‑scale transformer training. It signals a broader industry shift toward a multi‑vendor hardware ecosystem where cost, performance, and geopolitical resilience are balanced against software flexibility. Enterprises that adopt a pragmatic, data‑driven approach—leveraging GPUs for versatility and TPUs for efficiency—will not only achieve tangible savings but also position themselves to navigate the evolving AI hardware landscape with confidence.