AMD MI400X Major Launch: Performance Surpasses H100, Aiming for NVIDIA's Computing Power Throne
On July 29, 2026, at the AMD Data Center Conference in San Jose, AMD officially released its next-generation AI accelerator, the MI400X. According to official data, the MI400X uses advanced 3nm process and Chiplet architecture, integrating over 200 billion transistors, with FP8 computing power reaching 5.6 PFLOPS, more than 5 times that of NVIDIA H100 (1.0 PFLOPS). AMD CEO Lisa Su stated at the conference that the MI400X will enter mass production in Q1 2027, with initial customers including Microsoft, Meta, and Oracle.
After the announcement, AMD's stock rose 8.2% in after-hours trading, while NVIDIA's stock fell 3.5%, as the market fears NVIDIA's monopoly in AI computing chips is under threat. NVIDIA's previously planned Blackwell B200 chip is expected to launch in 2025 but faces challenges with heat dissipation and yield, potentially delaying mass production. AMD's MI400X directly competes with the B200 and outperforms it in performance metrics, undoubtedly intensifying the competitive landscape of the AI chip market.
Tech Giants' Capital Expenditure Soars: Microsoft Plans to Invest $80 Billion in 2027
On the same day, Microsoft announced in its earnings call that it will increase AI-related capex to $80 billion in 2027, nearly double the previously planned $45 billion. This figure far exceeds market expectations, mainly used to build new AI data centers and purchase GPUs and servers. Microsoft CEO Satya Nadella said: "The pace of AI adoption far exceeds our expectations, and we must significantly increase investment in computing infrastructure to support next-generation large models and cloud services."
In addition, Amazon AWS and Google Cloud have also recently raised their 2027 capex plans to $60 billion and $50 billion respectively. Meta plans to invest $40 billion in AI research and next-gen data centers. This means the total global investment of tech giants in AI computing power over the next two years will exceed $230 billion, more than double the 2025 level.
The surge in capex directly benefits the entire computing power supply chain. Server OEM leader Super Micro and Dell Technologies have announced capacity expansion to meet data center orders. Optical module suppliers Zhongji Innolight (overseas subsidiary) and Coherent have received a large number of 800G and 1.6T optical module orders. In addition, power infrastructure and liquid cooling solution providers are experiencing explosive growth.
Large Model Computing Demand Explodes: Training Cost of Trillion-Parameter Models Drops to Tens of Millions of Dollars
The launch of AMD MI400X and the expansion of tech giants' capex are driven by the exponential growth in computing demand for large model training. Companies like OpenAI, Google DeepMind, and Anthropic are developing models with over 10 trillion parameters. Currently, training once using H100 clusters costs hundreds of millions of dollars. AMD claims the MI400X can reduce training costs by over 80%, and combined with a more efficient interconnect architecture, is expected to bring the cost of training ultra-large models down to tens of millions of dollars. This trend will accelerate the adoption of AI applications. Inference also benefits: the MI400X's energy efficiency reaches 5 TFLOPS/W, far exceeding the H100's 3 TFLOPS/W, significantly reducing data center power consumption. According to IDC, the global AI server market will exceed $200 billion by 2028, with inference servers increasing from 35% in 2025 to over 50%.
Competitive Landscape Reshaped: How Long Can NVIDIA's Moat Hold?
For a long time, NVIDIA has dominated over 80% of the AI chip market with its CUDA ecosystem and solid hardware performance. Although AMD's MI400X leads in computing metrics, software ecosystem remains its biggest weakness. AMD is accelerating the promotion of the ROCm open-source platform and pushing native adaptation with mainstream frameworks like PyTorch and TensorFlow. In addition, Microsoft Azure has announced full support for the MI400X, helping it quickly enter the enterprise market.
On the other hand, NVIDIA is also actively countering. It is rumored that NVIDIA will launch the B300 chip in 2027, using 3nm process and a new architecture, with performance expected to triple that of the B200. However, analysts point out that the biggest challenge NVIDIA faces is not a single product, but customers' concerns about relying on a single supplier. Microsoft, Google, and Meta are all developing their own AI chips. In the long run, the competition in the computing power market will become more diverse.
Investment Insights: Seize the Structural Opportunities in the Computing Power Chain
For US stock investors, AMD's MI400X and the acceleration of tech giants' capex provide clear investment logic. First, the AI chip field will shift from "NVIDIA's dominance" to "duopoly" or even "multi-polar competition", with both AMD and NVIDIA benefiting from rapid market growth. Second, the data center infrastructure chain (servers, optical modules, liquid cooling, power) has stronger certainty and high order visibility. Finally, the continuous iteration of computing chips will drive cost reductions for AI application companies, such as AI SaaS, autonomous driving, and robotics sectors, which are expected to see profit inflection points.
It is worth noting that the current AI computing power sector valuation is at a historically high percentile. Investors should be alert to risks such as capex falling short of expectations or product delays. However, in the long run, the AI revolution is still in its early stages, and as core infrastructure, computing power's investment value cannot be ignored.
