Since I wrote my previous article on the AI Ecosystem on June 1, there have been interesting developments; all along expected lines, not surprising at all.

First, let’s look at the developments, and then we will discuss AI’s impact on IT Consulting. So, here is what happened:

The AI buildout was, until recently, mostly financed by hyperscalers’ cash flows and internal resources. Then there were all the circular financing deals across the ecosystem, including the neocloud GPU-backed debt. While it sounds smart to highlight these as issues, yes, optically it looks bad, but it's all legal, and the hyperscalers have enough balance sheet room to keep raising capital.

It is quite entertaining to discuss whether there is a bubble; however, building a portfolio based on a binary answer to that debate would be foolhardy. It is not about selling everything or getting leveraged to neck long. The answer is much more nuanced.

Given the risks, how much exposure do I want to have to a particular stock? Do I need to hedge? What multiple am I willing to assign to a stock? Or do I avoid a stock? I understand that this is common knowledge to most of you, but I still wanted to put it out there. So, what am I trying to figure out here?

Are we at an inflection point?

The risks, as I have highlighted previously:

If one pays close attention, the market has been pricing in some of these risks. It shows up in individual stocks. NVIDIA has been flat since November 2025. Oracle is suffering due to increasing leverage. Multiples remain compressed across the ecosystem even after the current parabolic rally. Essentially, the market is unwilling to assign higher multiples due to cyclicality and perceived risks.

With each passing day, the incoming news indicates the risks are compounding. The probability of a selloff in the AI space is increasing. That is a nice way of saying I am using Bayes’ Theorem.

In short, the new information we have:

I discussed the issues in Private Credit a few weeks ago; however, now we have the first evidence of a “big boy” being unable to raise capital.

In May, I made a dramatic pronouncement: OpenAI will be another WeWork for SoftBank. I do stuff like that to get conversations started. However, that statement wasn’t without merit.

OpenAI makes up ~30% of SoftBank's NAV. 70% of SoftBank’s NAV is ARM and OpenAI. The ARM stake is backing margin loans used to fund OpenAI. SoftBank tried to borrow $10 billion against the OpenAI stake to increase its stake in OpenAI. Banks said that’s too much. SoftBank cut it to $6 billion. Banks still said no. There are no prizes for guessing what SoftBank’s stock is doing.

While it’s hilarious in some ways, it is also mind-numbing. One can infer two things from this incident:

Hence, while the probability of financing running out across the ecosystem remains low in the short term, 2027 is potentially when it gets really interesting. Remember, the hyperscalers need to raise ~$1.5 trillion between now and the end of 2027.

TokenMaxxing is professionals and companies deliberately maximizing the consumption of AI tokens. It was a vanity productivity and learning metric used by clueless companies with no regard for ROI (Return on Investment).

Any new technology innovation brings out rational and irrational behavior. In the initial stages of the gold rush, the sane, rational people are chided and insulted. The same was happening with AI spend. Now that bills are showing up with no real benefits to show for the spend, reality is slowly setting in. I recently heard a new term making the rounds: OutcomeMaxxing.

Remember what I said about the tech industry in my previous article? They have a habit of rebranding old stuff and charging a premium for it. They couldn’t come around to saying they will focus on ROI. OutcomeMaxxing sounds so cool, doesn’t it?

One might wonder why people keep repeating the same mistakes across every cycle. It boils down to how humans have evolved. For tens of thousands of years, humans were hunter-gatherers. Every action was motivated by the most immediate need, whether it be food, shelter, or sex. The human brain evolved under that blueprint. Everything was: Now, now, now. It is only over the last few thousand years that long-term thinking has become a part of humanity. However, the brain structure still resembles the primal brain. The innate requirement for instant gratification remains ingrained.

Most of the time, this leads to not-so-well-thought-out irrational behavior. That is why people love gambling. That is why we have bubbles and crashes. That is why we have TokenMaxxing.

Finally, people are coming to terms with the essence of Goodhart's Law. It states: when a measure becomes a target, it ceases to be a good measure. Coined by economist Charles Goodhart in 1975, it warns that once a metric is used as a goal, people will optimize for that specific number, manipulating the system and losing sight of the original intent. And people are wondering why Tokenmaxxing was such a bad idea to begin with. Really?

In a way, the regime of irrationality is evolving towards a regime of rationality. This is a welcome and heartwarming change.

I have not hidden my disdain for Mark Zuckerberg over the last few years. Meta got away with a lot. It regularly restructured out of bad bets. Each of those was quite bullish for Meta. Well, why not? Wall Street loves restructurings. This time, however, the numbers are too big for a restructuring-fueled resurgence.

It Tokenmaxxed like companies with leaders who proved they are rookies rather than seasoned, experienced executives. The bigger damage this time? The morale of employees is destroyed.

So, Meta, like Amazon, Uber, and a plethora of others, burned tens of billions of dollars tokenmaxxing and has now instituted curbs on token usage. At least Mark admitted he made mistakes. However, it would not be surprising if there is a shareholder revolt at Meta sometime in the future. For that to happen, the stock has to crash first and stay depressed for a few years. In any event, time is not on Mark’s side.

Then there is SMCI. One of the things I highlighted in my May 2024 article about SMCI was their lack of capital to handle any potential large orders. I warned that they might have to dilute shareholders or raise more debt. Plus, their inventory and accounts receivable were ballooning. Take a peek at their balance sheet. Same situation now.

So, SMCI announced a plan to raise $7 billion, including $5B through public offerings and up to $2B through an at-the-market stock program. The stock fell ~7% after-hours when the news came out. The bullish narrative was that this was to fulfill ~$49 billion in AI-Server orders from 20 clients.

There is a big problem with that bullish thesis. I was like, this thing is going to tank hard tomorrow. And it did. Ended the day down 28%. So, what was the problem?

Inability to raise working capital financing to bridge short-term cash flow gaps through traditional banking channels and reliance on equity financing, essentially, selling long-term assets for working capital needs, is a big red flag. This implies SMCI couldn't pass the banks' underwriting checks, which brings into question the quality of orders.

Game, set, and match.

In any event, I have no interest in SMCI anymore, nor do I have any long-term opinions. I had my fun with them in 2024. I discussed them to highlight the following:

Frontier models for consumers are either free or subsidized. Consumers seem to be happy, and consumption is growing, albeit at a massive loss for frontier labs.

To capture the consumer market, frontier labs had to cover a large breadth of topics, essentially, the whole knowledge base of humanity, that they could get their hands on. The downside is that the models suffer from a false positive, ranging from 5% to 35%, depending on the quality of prompting and verification.

The growing risk for frontier labs is that consumer losses continue to rise. To allay this risk, frontier labs are attempting to capture the enterprise market. The problem is that frontier models are not built for the depth enterprises require. Enterprises have a low tolerance for errors, which frontier models aren’t ready to address yet due to their probabilistic nature. Enterprise solutions would also require more contextual data, targeted training, and guardrails. Most of which doesn’t exist.

While Anthropic talks a good enterprise game, all of the enterprise solutions currently in production are mostly Anthropic’s LLM running on Palantir (PLTR). People pontificating that Frontier Labs will somehow make PLTR redundant don’t understand what PLTR does technically. They don’t understand ontology, and that PLTR is like an embedded operating system, the orchestrator. LLMs don’t have these capabilities.

Why do you think the Frontier Labs need Forward Deployed Engineers or Solutions Consultants? They need these people to go into every enterprise, build the capacity, enhance capabilities, and provide the solution. That is easier said than done. Getting into a Fortune 500, building trust, and then being given the keys to mission-critical systems takes a lot of time. The faster path to enterprise adoption is for Frontier Labs to piggyback on consulting companies or companies like PLTR that are already embedded in the Enterprise IT ecosystem.

So, there, Palantir is going to be an AI winner. That said, I don’t like PLTR’s valuations, and their TAM (Total Addressable Market) has always been overhyped. I covered PLTR last year and said the same thing about their valuations and TAM.

To identify winners in the Enterprise AI space, one needs to understand the Enterprise IT ecosystem. A picture is worth a thousand words, so here is what those chasing Enterprise AI Dreams have to deal with.

The reality highlighted underneath the surface exists to varying degrees in Enterprises. Depending on the organization, there could be hundreds, if not thousands, of disconnected and disjointed systems and silos marred by inefficiencies and red tape. Some stuff is still being done using Excel sheets.

Depending on how an enterprise’s IT has evolved over the years, AI adoption could be a walk in the park or a nightmare. It all depends on the levels of integration between systems, whether they have data warehouses or data lakes, and the plethora of technologies from different vendors. Even then, a lot of custom work would be required.

Now, one may contend it can’t be that bad, or some may balk at the monster they are staring at. Well, I have news for them, this is absolutely true, and this is how it is.

Why do you think a company like Snowflake (SNOW) even exists? They shine at enabling organizations to store, manage, and analyze massive amounts of data in a single, unified system from siloed data sources. There, that is another AI winner for you. But, then again, I think it is overvalued.

PLTR’s ontology provides similar capabilities, along with it being actionable.

The biggest reality is that the systems of books and records at most of the large enterprises are still on Mainframes. Those are not going away anytime soon. It might come as a surprise, but the majority of daily transactions people are doing hit Mainframes. A very common daily transaction is a credit card transaction. Use a Visa or American Express; it gets processed on their Mainframe.

Some people, mostly San Francisco tech bros and DOGE, have been vocal about migrating mainframe-based systems to newer technologies. Best of luck with that. It will take at least a decade, if not more, and even then, the new system may not be as stable as the current ones. No CIO (Chief Information Officer) is going to take that level of risk.

There was a whole story about migrating the Social Security Administration’s systems off the mainframe. It was budgeted for, but was abandoned midway. As I said, there was a whole story, and quite an interesting one. I will leave it at that.

So, how are these systems still maintained and running? IT consulting companies help maintain and upgrade these systems.

Counterintuitive? Goes against the narrative? Let’s debunk the narrative.

The main selling point pushed by Frontier Labs and the AI community was that ~50% of white-collar jobs will become redundant. You, I, and the majority of people were not the intended audience for that message. The narrative was built and sold to VCs and private investors as the Total Addressable Market (TAM). When one looks at the math, it starts making sense.

Globally, corporates spend ~$40 trillion on white-collar labor. 50% of that is $20 trillion. When one looks at that TAM, it is easy to make a case for trillions of dollars of CAPEX.

Recent developments have proven that whole narrative to be wrong. The big question now is: what is the actual TAM for Frontier Labs now? Whatever futuristic stuff, like robotics, is thrown out there is not happening any time soon. It will happen eventually, but the economics of it don’t align with the current spend. The Frontier Labs will run out of money by then.

Setting that aside for a minute, there has been a perceptible change in messaging from the tech leaders. Now, the messaging is targeted towards productivity, and Agentic AI enabling Enterprises. Even Nvidia’s Jensen Huang has changed his tune on job losses. So, what gives?

What the San Francisco AI bros are realising is that they are, how to put it mildly, not liked outside of San Francisco. The problem with that is, who is going to buy their IPOs? Hence, the messaging change. But the question remains: what is the TAM? Certainly nowhere near $20 trillion. They are not getting that from enterprises, as I explained in my previous article. Well, global IT spend is $6.31 trillion. How much of that is AI going to get? Not enough to justify trillions of dollars in CAPEX. Actually, what they will get will not even cover the CAPEX.

Let me further show you why enterprise spending on AI will be much lower than what is anticipated.

We already know consumers are loss leaders for Frontier Labs. Then the main AI use case was coding. Reality shows it is not going to be as big as people expected. Plus, corporates require ownership of code and artifacts. The things that fly at startups and tech companies are a complete no-go at corporates. So, what is left? Agentic AI?

This is how the decision-making behind AI implementation at large corporations would go.

So, setting aside the fact that the Frontier Labs are pretty much cooked. We can see that, AI or not, custom work will continue to be required, and consultants will be required to enable AI implementation.

As IT consulting companies have been embedded in corporations for years, they know the ecosystem; without them, AI implementation at a large scale is close to impossible.

Now, coming back to #2. Can a lot of it be done with existing technologies? Yes. In short, cron jobs and loops. You can define workflows that need automating, custom code the logic, most of it would be required even in the case of using AI agents, and be done with it.

One can have PLTR as the operating engine, or enhance existing systems to be the operating engine, as they are already interacting with data. The pipes, connections, and integrations are already there. Plus, we have already discussed that LLMs are not built for stuff like this.

So, to do all of this, one will need IT consulting companies, mostly their existing partners. Whatever happens, IT consulting companies are not going anywhere. Most importantly, they will still be required to KTLO (Keep The Lights On). Plus, they have the trust and institutional knowledge that the LLM startups fronted by software engineers can’t even dream of competing with. Rather, the frontier labs will end up relying heavily on IT consulting companies. They actually already have partnerships with all of them.

Now, let’s take it a step further. Given that the economics of using frontier models don’t work, this is how the choice of or use of models will be broken down:

Did I not say the Frontier Labs are cooked?

That said, since we are on the topic of IT consulting companies, who are the winners?

Indian IT outsourcing companies not only face competition from AI models, which directly impact their cost arbitrage business model, but they also have to contend with competition from their own customers. Over 2,200 top companies by market capitalization from across the globe have opened Global Capability Centers (GCCs) in India. The offices are much more than back-office support work. The centers complement corporate functions, including R&D and IT. Nvidia has 30% of its global workforce in its GCC in India.

If I had to choose a possible winner among the Indian IT consulting companies, it would be Cognizant (CTSH). There are a few smaller ones, but they are listed in India, not in the US.

On a side note, Blackstone and Brookfield are making hay with their office REITs in India.

Finally, are we getting to an inflection point where there is a selloff in the AI space? I think we are close. Me talking about it on this blog and laying out the facts is one thing; it all becomes material when it goes mainstream, and there is a trigger event.

While people have not been able to get to the point where I am with Frontier Labs, and why they are cooked, Morgan Stanley has been very active recently in laying out exactly what I have been saying over the last month or so. Everything, to the dot. They even put out a bearish note on GEV. So, slowly, it is going mainstream.

The big one recently was that Citadel Securities put out a new macro note a few days ago titled "Tokenomics." Questioned the economics and viability of Frontier Labs and basically called out tokenmaxxing. They laid it out plainly:

As I like to say: ROI has entered the chat. If nothing else, we can look forward to the KOSPI and trillion dollar market cap companies trading like meme stocks for now. It is quite fascinating and scary at the same time.