---
title: "Why Amazon Says \"2028 Demand Is Striking\": From Labs AI to Enterprise AI Driving a Steeper and Longer Demand Curve"
type: "News"
locale: "en"
url: "https://longbridge.com/en/news/294620990.md"
description: "During the Q2 2026 earnings call, Amazon CEO Andy Jassy stated that the compute demand the company is currently seeing for 2028 is \"striking\" and cannot be met by existing capacity. He pointed out that the AI adoption curve exhibits a \"barbell\" effect: on one end are AI Labs consuming massive amounts of compute, and on the other are enterprise applications that have achieved real ROI (such as customer service automation). Together, these forces are driving steeper and longer demand growth"
datetime: "2026-08-03T00:39:29.000Z"
locales:
  - [zh-CN](https://longbridge.com/zh-CN/news/294620990.md)
  - [en](https://longbridge.com/en/news/294620990.md)
  - [zh-HK](https://longbridge.com/zh-HK/news/294620990.md)
---

# Why Amazon Says "2028 Demand Is Striking": From Labs AI to Enterprise AI Driving a Steeper and Longer Demand Curve

Regarding the intensity of market demand for compute power, many people will likely be deeply impressed by a statement made by Amazon CEO Andy Jassy during the latest earnings conference call (Q2 2026 Earnings Call):

In fact, the demand we already have for 2028 is striking. In fact, the demand we are already seeing for 2028 is astonishing.

The word "striking" literally means shocking or extraordinary in Chinese.

Clearly, this is a point worth serious attention. More importantly, after stating that "even at this scale ($220 billion in CapEx), we will still not be able to meet all existing demand in 2026. I believe this situation will continue in 2027. In fact, the demand we are already seeing for 2028 is striking," Jassy did not attribute this demand solely to AI Labs.

Instead, he proposed a compelling "AI demand barbell":

I think it's actually kind of useful to look at at least our view of what we see in the demand and adoption curve right now, which is we see this adoption curve in AI right now is very barbellled. I think it is quite useful to look at our current view of the demand and adoption curves; specifically, the AI adoption curve currently exhibits a strong "barbell" effect.

On one end are AI Labs, which are consuming massive amounts of compute power; on the other end are enterprises that are already deriving real ROI from AI, such as through customer service automation, business process automation, and fraud detection.

The two ends of this "barbell" are heavy, while the middle is light. The light middle section represents the large volume of production workloads and processes that currently exist in enterprises. Jassy believes that most of this segment has not yet widely adopted inference, but significant changes will occur in the future. He directly offered his judgment during the conference call:

The largest segment in the end will precisely be the middle part of the "barbell."

This is a very important key signal indicating that the logic behind AI infrastructure investment is changing. Therefore, when explaining why compute demand is so tight, Jassy used the following statement:

Remember, enterprises are still very early in using inference at scale in their current production applications. Please remember that enterprises are still in a very early stage of using inference at scale in their existing production applications.

Because in past discussions about who pays for AI CapEx, the market mainly saw Frontier Labs like OpenAI, Anthropic, Google, Meta, and xAI frantically buying compute power.

But this time, Jassy told the market:

This is only the first layer. The truly massive demand has not yet fully emerged.

Why? Because the denominator is too large.

There are only dozens of AI Labs, but there are millions or even tens of millions of enterprises globally. Each enterprise has hundreds, thousands, or even tens of thousands of production workflows.

This is also why Jassy dared to say this time:

AWS could potentially become a $1 trillion revenue business in the long term.

(I)

Google and Microsoft Are Also Proving Enterprise AI

Meanwhile, Google's latest quarterly earnings report has actually proven that Enterprise AI is kicking off.

This signal is also very strong. Google stated:

Nearly 90% of Fortune 100 companies are already using Gemini Enterprise, and enterprises are further increasing their usage, with over 2,000 companies processing more than 100 billion tokens in the past year...

"PepsiCo uses our AI and analytics solutions, Intel uses our solutions to simplify core processes, HSBC uses our solutions for wealth management, Bell Canada uses our solutions to enhance customer engagement, Macy's uses our solutions to optimize the e-commerce experience, and SIGNAL IDUNA uses our solutions for knowledge management."

When Google puts these numbers and cases together, especially these vivid Enterprise AI application cases, it shows that Enterprise AI has started to move from:

"Enterprises trying out AI" to "Enterprises continuously consuming tokens."

Microsoft's latest earnings report further proves that Enterprise AI is moving from "Seats" to "Workloads." Azure's commercial RPO for this quarter reached $678 billion, an 84% year-over-year increase. However, Microsoft particularly emphasized:

All sequential commercial RPO growth was driven by commitments from customers outside of frontier model companies. All sequential growth in commercial remaining performance obligations was driven by commitments from customers other than frontier model companies (i.e., AI Labs).

In other words, the sequential RPO growth for Azure in the just-concluded quarter reflects:

AI infrastructure demand is spreading from AI Labs to Enterprises.

This is a very critical change, indicating that Enterprise AI is undergoing an evolution from Seat → Agent → Workflow → Production Workload.

(II)

Labs AI: RSI Will Steepen the Demand Curve Again

For Frontier Labs, the most important change currently is Recursive Self-Improvement (RSI). Anthropic's blog post "When AI builds itself" on June 4 this year was the first signal of the RSI process. The article stated:

-   Within Anthropic, an increasing amount of AI development work is being handled by AI systems themselves;
-   As of May 2026, Claude wrote over 80% of the code merged into the codebase, compared to single-digit percentages before the release of Claude Code in early 2025;
-   In tasks such as "optimizing experiments for given objectives," model performance has reached superhuman levels, accelerating from approximately 3x to about 52x.

Anthropic co-founder Jack Clark even publicly stated that the probability of achieving fully autonomous RSI by the end of 2028 is 60%.

On the OpenAI side, the recently released GPT-5.6 series, Sol (flagship), demonstrated a case: based on an incomplete prompt, it autonomously selected training configurations, GPUs, and post-training tasks to generate a smaller Luna model, estimated to save two senior researchers about two weeks of time.

Then last week, we saw news that Lilian Weng returned to OpenAI to lead RSI research. She had just published an in-depth long article discussing RSI, titled "Harness Engineering for Self-Improvement," on her personal blog in early July.

The importance of RSI for compute demand lies in:

When AI begins to participate in building better AI, compute demand will no longer be just "to meet user needs."

This will form a potential positive feedback loop:

Stronger models → Stronger Coding/Research Agents → Higher AI R&D efficiency → Faster experimentation speed → More training → Stronger models → Higher compute demand.

The entire process repeats itself, turning the training process into: continuous training, where training never stops.

Taking a step further:

AI writes training code → AI designs experiments → AI runs experiments → AI analyzes results → AI generates new training plans → AI participates in post-training → New models are born.

If this closed loop becomes increasingly complete, then compute demand will expand from:

"Serving AI users" to "AI consuming compute to improve itself." This is a completely different source of demand, and theoretically, there is no natural upper limit to this demand for compute power.

Therefore, a major shift in perception has occurred:

Coding Agents may not just be terminal applications, but rather the first door for AI to enter the AI R&D system. Consequently, the Demand Curve for Labs AI will likely steepen again.

Thus, it is impossible to simply extrapolate 2028 compute demand based on today's user numbers for OpenAI and Anthropic.

(III)

AI Has Entered the Phase of "Self-Reinforcement + Comprehensive Penetration"

Therefore, we see that AI compute demand is forming two distinct but mutually reinforcing growth engines.

-   One is Labs AI: Driven by Frontier Labs, it is shifting from Scaling to more powerful Reasoning, Agents, and Recursive Self-Improvement (RSI). Its characteristics are extremely high compute demand and a very steep growth slope.
-   The other is Enterprise AI: AI is beginning to enter real enterprise production systems, gradually penetrating all enterprise production workloads from partial scenarios such as customer service, coding, finance, and risk control. Its characteristic is not a short-term burst, but rather continuously increasing penetration rates, rising usage frequency, and extreme longevity.

For Enterprise AI, there is also a very important dual growth factor:

Penetration Rate × Compute Volume per Workload.

This is what makes it truly formidable. Can you understand the multiplication above? This is far more important than "whether AI has the next Coding application!"

The narrative that "AI's largest commercial application currently is Coding, and if there is no new trillion-dollar application after Coding, then AI CapEx may peak in 2028" has a fundamental logical flaw:

It interprets AI demand as "searching for a new super App." Enterprise AI may well take a different path, not by birthing an application larger than Coding, but by letting AI enter all existing applications and workflows today.

Therefore, Jassy's statement that "2028 demand is striking" should be understood as follows:

Labs AI + Enterprise AI

Ultimately, we will see a very rare industrial structure:

Upstream, RSI continuously pushes AI capabilities and compute demand upward; downstream, Enterprise AI continuously drives up AI penetration rates. One is responsible for making the curve steeper; the other is responsible for making the curve longer.

Risk Warning and Disclaimer

The market carries risks, and investment requires caution. This article does not constitute personal investment advice, nor does it consider the specific investment goals, financial status, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investment based on this content is at your own risk.

### Related Stocks

- [AMZN.US](https://longbridge.com/en/quote/AMZN.US.md)
- [AMZU.US](https://longbridge.com/en/quote/AMZU.US.md)
- [AMZD.US](https://longbridge.com/en/quote/AMZD.US.md)
- [AMZW.US](https://longbridge.com/en/quote/AMZW.US.md)
- [AMZZ.US](https://longbridge.com/en/quote/AMZZ.US.md)
- [AMZO.US](https://longbridge.com/en/quote/AMZO.US.md)
- [AMZY.US](https://longbridge.com/en/quote/AMZY.US.md)
- [AZYY.US](https://longbridge.com/en/quote/AZYY.US.md)
- [AMZP.US](https://longbridge.com/en/quote/AMZP.US.md)
- [MAGX.US](https://longbridge.com/en/quote/MAGX.US.md)
- [OpenAI.NA](https://longbridge.com/en/quote/OpenAI.NA.md)
- [GOOGL.US](https://longbridge.com/en/quote/GOOGL.US.md)
- [GOOG.US](https://longbridge.com/en/quote/GOOG.US.md)
- [META.US](https://longbridge.com/en/quote/META.US.md)
- [MSFT.US](https://longbridge.com/en/quote/MSFT.US.md)
- [PEP.US](https://longbridge.com/en/quote/PEP.US.md)
- [INTC.US](https://longbridge.com/en/quote/INTC.US.md)
- [HSBC.US](https://longbridge.com/en/quote/HSBC.US.md)
- [00005.HK](https://longbridge.com/en/quote/00005.HK.md)
- [BCE.US](https://longbridge.com/en/quote/BCE.US.md)
- [M.US](https://longbridge.com/en/quote/M.US.md)
- [DTIW.SG](https://longbridge.com/en/quote/DTIW.SG.md)
- [HSBA.UK](https://longbridge.com/en/quote/HSBA.UK.md)

## Related News & Research

- [Amazon, Walmart AI detect 'made in USA' fraud but do not flag it, study says](https://longbridge.com/en/news/294339580.md)
- [Discover top deals with AI-powered shopping this Great Freedom Sale on Amazon.in starting 7th August 2026](https://longbridge.com/en/news/294095182.md)
- [Meta's future AI data center lease obligations swell to over a quarter trillion dollars](https://longbridge.com/en/news/294389887.md)
- [BUZZ-Street View: Meta's AI push to weigh on free cash flow as investors seek returns](https://longbridge.com/en/news/294329489.md)
- [Resolute Advisors LLC Increases Holdings in Amazon.com, Inc. $AMZN](https://longbridge.com/en/news/294790842.md)