Last Resort?

Note: This article may at times feel meandering, and the tone may come across as alarmist. But it points to some important developments that can serve as a canary in the coal mine.

Share

Recent weeks have surfaced several developments in the AI ecosystem that, taken together, may offer important signals about the evolving trajectory of this current epoch in the AI journey. In this post, I examine these seemingly disparate developments, which share a set of common and critical underlying themes:

1. OpenAI introduced ads. Unsurprising to those who saw this coming (as I noted in an earlier post). Yet the CEO had previously claimed that ads would be a last resort for the company. Specifically, he stated that:

   “Ads plus AI is sort of uniquely unsettling to me…I kind of think of ads as a last resort for us for a business model”

   Following the OpenAI CEO’s own admission, it appears the company has now turned to what it once described as a last resort. I’ve pointed out many times, and explicitly, that one of the central challenges of the GenAI narrative is the absence of a viable business model. OpenAI’s move suggests an acknowledgment of that problem, even if not stated outright.

   And it’s not just the implicit admission. Reports indicate that the recently announced $100B deal with Nvidia is far from certain. Nvidia is reportedly hesitant to invest, with one source noting that its CEO “has also privately criticized what he has described as a lack of discipline in OpenAI’s business approach,” alongside concerns about competitive dynamics.

   Two points are worth underscoring both of which many experts have been raising for a long time. First, viable business models for GenAI are non‑trivial in most cases, despite the prevailing narratives. Second, the supposed “moats” derived from having an edge in foundational models rarely hold for long, if they exist at all. More than a year ago, I noted that foundational models were bound to converge, and that the moat would not lie in the models themselves despite the industry’s habit of treating each incremental benchmark “beat” as a breakthrough.

   I also highlighted the massive bet on scaling, among other factors, and pointed out that this too is not necessarily differentiating. In fact, it risks becoming an economic liability.

   Now, from OpenAI’s perspective the logic can be that they can’t offer service for free and organically people aren’t willing to pay. And hence paid tier or ads (plus other earlier decisions such as the introduction of and tolerance for erotica).

   The flip side raises several important questions with wide‑ranging implications. Does the service meaningfully add value for users, or is it primarily a polished plaything? If it were truly delivering substantial value, one would expect far greater willingness to pay. That clearly isn’t happening, and nothing suggests the paid subscriber base will improve under the current model. This points to two possibilities:

   1. users don’t rely on it heavily, meaning the free tier is sufficient; or

   2. users don’t perceive enough value to justify paying.

   In either case, the concerns extend well beyond OpenAI and speak directly to the broader value‑addition claims of GenAI. Is ChatGPT an impressive, polished offering that ultimately delivers only incremental value? Has there been an internal realization that the competitive moat is eroding far faster than expected (something many observers have been warning about for a long time) prompting these last‑resort measures following the earlier “code red” posture?

   More broadly, the central question remains whether current capabilities provide value commensurate with the scale of investment. Or has the narrative of doubling down on spending, justified by promises of future exponential growth, finally beginning to collide with the reality that either the scale of the promise or the timeline no longer makes sense? If so, the industry will face increasing difficulty in realizing the ROI that GenAI has been expected to deliver. And if that happens, the economics will only become shakier as investment commitments continue to rise and these bets expand further.

   As an aside, the tech advertising model is inherently inflationary. The cuts that ad platforms take from businesses (for example, when a retailer like Walmart sells directly through the OpenAI interface) ultimately get passed on to consumers. And not just to people using say ChatGPT. These costs diffuse across the entire market, often giving businesses convenient cover to raise margins. Over time, this dynamic risks making markets more concentrated and more monopolistic.

   The deeper question is whether these higher prices come with any comparable value‑add delivered through ChatGPT or other GenAI interfaces. For users, what is the real cost of using ChatGPT? It is likely far higher than the so‑called “free” tier suggests.

   In parallel, many companies are reducing headcount, and wages are already shrinking in anticipation of “imminent” GenAI‑driven value realization. In effect, firms are willing to sacrifice not only talent but also the fundamental structures that have historically enabled businesses to operate and sustain themselves, all in expectation of future AI impact. Regardless, the labor market and by extension the overall consumer base can become increasingly constrained in its ability to afford goods and services.

   None of these trajectories lead anywhere good. They point toward the risk of mass social unrest, an increasingly exploitative economic environment, the expansion of the “surveillance economy”, or some combination of all three.

2. Earnings from Microsoft and Meta: Recent earnings calls from the two companies produced diametrically opposite market reactions despite both reporting decent numbers. Microsoft faced significant pressure (its stock dropped nearly 10%), while Meta received a positive response. The general consensus is that Microsoft, even with improved cloud sales, highlighted the risks of relying on less reliable GenAI‑related revenue promises and the lack of clear indicators that value is actually being realized from GenAI applications (as opposed to merely providing the “shovels”). Meta, on the other hand, appeared to demonstrate tangible ad‑revenue gains driven by AI across its platforms.

   These developments raise interesting questions. When combined with the market reactions, the corporate push for value realization, and the growing trend of supposedly innovative companies reverting to the advertising model (not just OpenAI, but also Perplexity and even the hyperscalers) it becomes increasingly intriguing to ask whether the primary “value” of GenAI ultimately collapses back into advertising.

   Moreover, when contextualized against personal experience (at least anecdotally), it is not clear whether:

   • the content generated on Meta’s platforms alone is actually driving the ad‑revenue gains, or whether automated accounts and agents are also functioning as “consumers” in these systems. In other words, are AI bots effectively accounting for both sides of the equation?

   • there is meaningful, scalable value (realizable ROI) predominantly in loosely governed applications with limited accountability, such as Instagram posts, YouTube videos, and other social‑media‑driven content ecosystems.

   • and finally, whether Meta even needed GenAI investments at the scale it pursued in order to achieve the economic outcomes it now claims. From the outside, this does not appear to be something that required tens, or potentially hundreds, of billions of dollars of investment. One would then assume that these investments are aimed at significant future opportunities. How the value-realization will manifest is an open questions. And hence, the question of whether the business model is viable relative to the scale of investment applies even to Meta.

3. OpenAI continues to lose money, and its lack of a viable business model is now creating spillover risk for all associated parties. Investors are increasingly uneasy, and that anxiety was reflected in Microsoft’s post‑earnings stock drop (~10%) given that roughly 45% of its projected future Azure revenue is tied to OpenAI. It’s no surprise, then, that many hyperscalers are rushing to sign additional ad‑related or model‑access deals with other players in the space, including Anthropic and Perplexity (I’ve discussed the nature of these deals and the evolving economic risks in more detail elsewhere). Whether this represents genuine diversification or simply an attempt to hedge market perception and derisking the perception of dependence on OpenAI remains to be seen.

4. Companies are still waiting for meaningful ROI from GenAI. Adoption, predictably, is proving difficult, especially for safety‑critical or mission‑critical businesses. Bank of America is a good example of this struggle. As Business Insider reported, internal email threads reveal that actual adoption is far behind public perception (it may have not even started). This is not surprising: selecting infrastructure, choosing models, or assembling an AI stack is the easy part, particularly when done without context or expertise. Real adoption requires understanding transformation challenges (I have covered these aspects for GenAI earlier). These challenges go far beyond an executive declaring the company “AI‑first” for PR purposes.

   But didn’t Bank of America already claim massive impact and global adoption last year? It did. So what explains the discrepancy? Was the earlier announcement premature, or is the current iteration incomplete? To be fair, last year’s announcement likely referred to the deployment of customer‑service agents, a narrow form of workflow automation, not a fundamental transformation. Even if some impact was visible, it was not transformative. It did, however, provide the PR boost that many companies have been chasing.

   I won’t dive into the details of AI adoption or the increasingly performative nature of AI‑related IP claims. The patent game has become its own spectacle, with many filings lacking novelty or defensibility. For those who understand the AI ecosystem, it is clear that establishing and enforcing broad AI patents is extremely difficult. Bank of America’s claim of more than 1,200 AI patents does not necessarily reflect depth or strength in AI; it certainly does not align with the evident lack of platform understanding.

   This is not about singling out any one company. Bank of America simply serves as a representative example, one of many I’ve encountered over the years. Opportunistic misstatements are widespread across industries, enabled by abysmal reporting standards and norms. We’ve seen similar cycles of overclaiming and subsequent reversal in cases like Klarna, Taco Bell, and others. Alas, we live in a weird world:)

In essence, what is becoming increasingly evident is the absence of viable business models and the inherent difficulty of adopting GenAI at scale within established organizations. In the short term, however, companies are finding creative ways to monetize the narrative, even as meaningful ROI remains elusive. While there are exceptions and genuinely thoughtful efforts, the following generalization does not seem far off:

Despite all the promises of using GenAI for the betterment of the world, the most promising avenue appears, once again, to be the advertisement model powered by the “surveillance economy” (in its most benign interpretation). Most “innovation” on both sides of this equation, capturing user behaviors and, developing and serving monetizable content, is precisely where GenAI is being deployed most aggressively.

Of course, one can point to the rapidly expanding list of GenAI applications in education, counseling, medical care, law, the information ecosystem, and beyond. Yet we still lack evidence of reliability in any of these domains. There are numerous instances of harmful (sometimes even fatal) outcomes that illustrate the risks created by this combination of unreliable systems and rapid adoption.

On one hand, the capabilities and underlying systems are clearly not mature enough for seamless deployment in such critical applications. On the other hand, despite the fast (and often reckless) push toward adoption, the actual success metrics are neither understood nor defined, let alone measured. We urgently need a much broader debate about the objectives of GenAI systems and the objectives of the application domains themselves, especially as both evolve in rapidly changing environments.

As actionable items: For AI transformation, businesses need to prioritize meaningful efforts rather than chasing hype. And for genuine social betterment, we collectively need to pivot toward a more robust framework centered on social outcomes and priorities as the anchor for AI development and integration.