Is the Future of AI a Bubble or a Revolution? — The Risks and Possibilities Behind Massive Investments

Introduction: The Controversy Title is "Inference"

The Nextbigfuture article "AI Inference Boom or AI Ponzi Bubble" published on October 10, 2025, organizes the recent series of massive deals involving OpenAI×NVIDIA×AMD under the framing of whether they represent a real business revolving around "Inference" or a bubble inflated by "circular funding" [The article is dated October 10, Pacific Time]. The article acknowledges the cycle of investment → chip purchase → stock value increase → reinvestment but concludes that it is not a "Ponzi" scheme as long as there is an accumulation of tangible assets like new data centers/chip shipments/inference billing. It cites the fact that NVIDIA and AMD have contract structures that ensure sales and profits and that demand from OpenAI is ongoing. NextBigFuture.com

Facts: The "Two Wedges" of 10GW and 6GW

First, NVIDIA×OpenAI. The two companies announced plans for OpenAI to deploy up to $100 billion in investment capacity and at least 10GW of NVIDIA systems in phases. The first 1GW is set for late 2026, with the platform being Vera Rubin. Investments will be executed in phases corresponding to the actual deployment of each GW. This is confirmed by official announcements and Reuters. NVIDIA NewsroomOpenAI

Next, AMD×OpenAI. In addition to supplying 6GW of GPUs, AMD will grant OpenAI warrants to acquire 160 million shares at $0.01 each (equivalent to about 10%) in phases. The rights are tied to milestones such as achieving delivery of 1GW and AMD's stock price reaching $600, structured as quantity and price milestones, with vesting in tranches rather than immediate dilution (as reported by AMD's IR, OpenAI, and technical media). Advanced Micro Devices, Inc.

This **"10GW + 6GW" forms the core of what Nextbigfuture describes as "circularity." NVIDIA invests in OpenAI, which then uses that capital to purchase NVIDIA/AMD chips**. AMD enhances OpenAI's incentives with stock warrants, and OpenAI's success reflects back on AMD's stock price increase (OpenAI's unrealized gains). While the article points out this self-driving loop, it argues that as long as the real economy of "sales, construction, and operation" is active, it does not meet the Ponzi definition (a "dividend-less chain" distributing to old investors with new funds). NextBigFuture.com

Counterarguments and Concerns: How External Eyes View "Circularity"

On the other hand, external media are cautious. While Reuters reports the **$100 billion investment capacity as a fact, it touches on the possibility that the structure linking supply and investment could raise concerns about antitrust and circular revenue. Bloomberg features these "webs of interdependence" as a "chain of circular deals," highlighting voices warning of market overheating. Furthermore, Breakingviews editorializes that there is "no such thing as a 'good bubble'**," emphasizing the recession risks of AI investments reliant on debt and expensive specialized hardware. It points out the current situation where there is a lack of macro safety nets. Reuters

Social Media Sentiment: A Polarized Timeline

On X (formerly Twitter) and Reddit, opinions range from the strong assertion that "AI is a Ponzi" to the defense that it is an "industrial policy with aligned investment and supply," resulting in a split evaluation.

• Optimists argue that the "$100 billion phased investment is linked to the progress of GW deployment and represents a funding cycle backed by actual demand." Indeed, posts on X reflect the understanding that **"funds are released per GW"**. X (formerly Twitter)

• Skeptics view the cycle of "investment → demand recognition → sales → stock price → reinvestment" as "artificially inflated demand," criticizing AMD warrants as a scheme to "return 'rebates' on sales in the form of stock." On Reddit, the term **"AI Ponzi"** frequently appears. Reddit

Both sides ultimately focus on the operating rate and unit economics post-2026. The **launch of the first 1GW (late 2026)** is likely to be a leading indicator. NVIDIA Newsroom

The Reality of the Inference Business: The Key is "Operating Rate × Power × Customer Mix"

"Inference" is a revenue model where, although the query unit price is lower than model training, high gross profit is earned through a large number of uses, functioning as a utility-type model. Nextbigfuture emphasizes the scalability of inference, evaluating it as a "venture-like bootstrap" of "infrastructure investment → inference sales → reinvestment." However, the potential for declining inference unit prices, rising power costs, and memory/storage constraints to pressure profitability remains. NextBigFuture.com

Furthermore, the risk of customer concentration is not easily visible. The dependency on top customers on the GPU supply side and the bias towards large workloads on the inference side could lead to operating rate shocks during economic fluctuations. The market views this as the boundary between **"circularity" and "sustained demand". In past IT bubbles, advance capital investment and sharp declines in operating rates triggered collapses. Breakingviews' statement that "there is no good bubble" is rooted in this **historical memory**. Reuters