Over the past month, the volatility of technology companies has increased rapidly, undergoing a process from low volatility to high volatility (earnings season), then to extreme volatility (this week), and gradually returning to high volatility (changes in expectations).
From a sentiment perspective, "panic" has begun to subside, yet high skepticism remains (regarding AI-related performance, actual demand, and revenue forecasts). Therefore, while a forward-looking perspective at this juncture might not be entirely "calm" (hence the quotation marks in the title), a synthesis of various information points clearly hints at potential changes in the coming period.
First, the critical information:
Personnel tremors at OpenAI: Greg Brockman is taking a long-term leave, using the sophisticated term "Sabbatical." Additionally, several core figures have reportedly joined Anthropic AI (the creator of the Claude model). Since OpenAI failed to release GPT-5 (GPT-Next) in June as expected and delayed the rollout of the Advanced Voice Model, market expectations have been declining (with the release of GPT-4o-mini, the OpenAI "halo" is fading). As a primary bellwether, OpenAI's series of below-expectation performances and internal instability continue to significantly impact market sentiment.
Rumored delays for Nvidia's next-gen GPUs (Blackwell architecture B100/B200/GB200 series): According to an analysis by SemiAnalysis, a delay is largely confirmed, with expectations pushed back by three months—from Q4 this year to early next year. This undoubtedly negatively impacts the training of larger models and affects the timeline for Nvidia's subsequent products. Alongside OpenAI, this news has severely dampened market sentiment.
Earnings: (1) Performance for Microsoft, Google, and Amazon in the cloud sector has generally been solid. While Microsoft Azure's growth slowed slightly (mostly due to hardware supply issues), Google Cloud and Amazon Web Services (AWS) exceeded growth expectations (though AWS growth couldn't entirely mask revenue concerns, and Intel is rapidly falling behind). (2) Super Micro Computer's performance fell below expectations. (3) IBM’s performance in AI SaaS significantly beat market expectations (ignored by many, IBM's results show that enterprise-facing AI is profitable). (4) Apple's recovery speed exceeded expectations.
Declining model inference costs: OpenAI continues to lower API call fees, with Gemini and Claude following suit.
In summary, we can discern the following threads of future change:
The focus will gradually shift from the models themselves to scenarios and business implementation: Although OpenAI underperformed, Meta's open-weight model Llama-3.1 has reached the level of GPT-4 and Claude 3.5. This means concerns regarding data privacy and usage costs are diminishing. Looking at the strategic roadmaps of large corporations across industries, digital and AI transformation is undoubtedly the top priority, with many having clear plans to increase capital, labor, and infrastructure investment in this field.
The hype for pure computing power is fading, while service growth is becoming more certain: With the rapid increase in high-quality model availability, maturing model distillation (smaller models performing better), and iterating inference technologies, demand from enterprises and individuals is shifting from pure super-servers (represented by Nvidia GPU servers) to cloud services (which offer lower comprehensive costs and better technical support). Beyond cloud servers, demand is rising for SaaS (or MaaS, Model as a Service), as evidenced by the performance of both the public IBM and the private Databricks.
Edge-side implementation carries increasing expectations for consumer applications: Optimism for AI phones and AI PCs exists because the sales base is large enough and the industrial chain is long enough; any replacement cycle can drive performance. However, since this round of AI changes interaction methods, the eventual carriers are unlikely to be limited to phones or PCs. Smart cars, humanoid robots, and wearable hardware offer greater imaginative potential, though they need time. Yet, the core competitive element of "technology first" remains, as does the model-driven monetization via large-scale infrastructure services and All-in-One structures (Wintel is unlikely to succeed at AI PCs).
Summary in one sentence: Cloud services and the edge-side ecosystem.
As we enter autumn and vacationing workers return to their posts, I don't expect high market volatility or AI skepticism to settle significantly in the short term. However, we can still anticipate more "excitement" this fall:
At the model level, OpenAI's Advanced Voice Model will gradually roll out to all paid users, and SearchGPT will also begin its rollout, rapidly improving the substance of "intelligent assistants." Claude 3.5 Opus (the largest model) will be released—if Sonnet is already considered the best by most, expectations for Opus are even higher. Gemini will see upgrades, and both Meta (Llama) and Mistral will remain active. Spring and autumn are peak periods for model releases.
The release of iPhone 16 featuring Apple Intelligence: Despite constant skepticism, Apple typically achieves the highest level of refinement in every major product form. If Apple cannot succeed here, it likely means other products will fare even worse.
Beneath the surface: An increasing number of enterprises and individuals are accelerating the restructuring of their businesses and workflows using models, increasing the proportion of AI applications in daily work, study, and life. I cannot predict the specific moment "quantity leads to quality," but I know it is coming soon—very soon.