Em Richard

Nvidia: 6 Business Decisions on the Road to $6 Trillion

Em Richard — Sun, 11 Oct 2026 20:52:12 GMT

Prologue: a bedroom in Toronto

A few days ago, I read a post from a well-known community manager. Once again: Nvidia, $6 trillion. Once again, the big number, the VC jargon, the hype. And not a word on how the company actually got there.

So I went looking for the story behind the number. What I found is the opposite of the hype. The largest AI company in the world, the largest market cap on the planet, was built on human instinct, human innovation and human intent.

It starts in a bedroom.

Toronto, 2012. A family house like any other. Upstairs, in the bedroom where he grew up, a graduate student named Alex Krizhevsky leaves a computer running day and night.

Inside the case, two graphics cards. GeForce GTX 580, the kind teenagers buy to play games, about $500 each. For five or six days, they grind through more than a million images, again and again.

On September 30, the results of ImageNet come out. It is the world’s reference contest for image recognition, and the big labs compete in it. Krizhevsky’s model, built with his fellow student Ilya Sutskever and their professor Geoffrey Hinton, wins by a wide margin. The paper will be cited more than 172,000 times.

That autumn, modern AI was born in a bedroom. On a machine built for video games.

Ask how that was possible, and you pull a thread. It runs back through a Stanford lab, a letter mailed to Taiwan, a humbling flight to Japan and a booth in a San Jose diner. It runs forward to a company worth close to $6 trillion.

This is the story of that thread, told the way a seller reads a company. The bets, the offer, the buyers, the money and the timing.

Chapter 1: six months to live

The thread starts in 1993, at a Denny’s on the corner of Capitol and Berryessa in East San Jose. Jensen Huang knows the place. He washed dishes at Denny’s, bused tables, waited on customers. He likes the bottomless coffee.

In a booth, with two engineers from Sun Microsystems, Chris Malachowsky and Curtis Priem, he decides to start a company. The idea: chips that bring 3D graphics to ordinary PCs. The market for it? Sequoia’s Mark Stevens, who would join the board, says any research firm would have sized it at zero. They start with $40,000 in the bank, then raise $20 million from Sequoia, Sutter Hill and others.

Their first chip, the NV1, ships in 1995. It fails. Nvidia had chosen its own way of drawing 3D shapes, and then Microsoft set the standard for PC games and chose another. Brilliant engineering, wrong side of history. Huang would later call the chip technically poor.

Worse, Sega had hired Nvidia to build the next chip for its console, on the same flawed design. Nvidia had cut half its staff. The cash was running out.

So in 1996, Huang gets on a plane to Japan. Picture the scene. A young CEO walks into the office of his most important client. He tells them the product is wrong. He tells them to cancel the contract. And he tells them that Nvidia still needs the last $5 million payment, or it will not survive.

Every sales instinct screams not to do this. You protect the deal. You manage the message. You buy time.

Huang told the truth. And Sega’s American CEO, Shoichiro Irimajiri, convinced his company to pay and to invest. Huang later said that generosity gave Nvidia six months to live.

Six months. Nvidia put everything into one chip, the RIVA 128, built in record time and launched in 1997. A million units in four months. Twenty percent of the PC graphics market. The first quarterly profit. Sega would later sell its Nvidia shares for $15 million. Three times its money, for the price of believing a man who admitted he was wrong.

The first lesson of this story is the oldest one in sales. When you are truthful and direct, even with bad news, you give the client a reason to bet on you.

Chapter 2: a letter to Hsinchu

To understand the second bet, you need to know what a chip is.

A chip is a slab of silicon carrying billions of microscopic switches, the transistors. Each one is either on or off, 1 or 0. Think of a city with billions of traffic lights. Designing a chip means drawing the city: where the roads go, how the lights are timed, so data flows to the right place, fast. Manufacturing it means actually building that city, with roads thousands of times thinner than a hair. Only a handful of companies on Earth can do the second part.

Nvidia knew how to draw cities. It had no way to build them. A chip factory costs billions. Nvidia had about 50 to 60 employees.

The answer was in Taiwan. TSMC, founded by Morris Chang, did one thing only: build chips designed by others. Nvidia contacted TSMC’s office in San Jose. Nobody took the small startup seriously.

So in 1997, Huang did what founders do when the front door is closed. He wrote to the top. A letter, by regular mail, to Morris Chang himself. It reached Chang in Hsinchu, Taiwan. A small company, it said, with some very promising chips. Would he please call.

The letter itself has never been published. What we have is Chang’s memory of it, in his memoir and in a rare interview. He remembers being curious, and a little irritated that his own team had ignored a customer. On a trip to California, he picked up the phone and dialed the number printed on the letterhead.

Huang answered himself. Behind him, a noisy room, a heated discussion. Chang introduced himself. And Huang shouted to his team: “Quiet, Morris Chang is calling me”.

At their first meeting, Huang made his pitch. Nvidia is in financial trouble. But this chip will save the company, and make us one of your biggest customers. Chang found it a bold statement. Within two to three years, Nvidia was one of TSMC’s top five clients. In 1998, when Nvidia was short of hands, TSMC sent two of its production planners to live inside Nvidia for a month.

That letter solved the hardest equation in the business: how to build the most complex objects on Earth without owning a single factory. It set Nvidia’s model for thirty years. Nvidia draws. TSMC builds. Every dollar Nvidia raised could go into design, software and people.

Sixteen years later, in 2013, Chang tried to hire Huang as TSMC’s CEO. Huang said no within ten minutes. He already had a job.

Chapter 3: developers, developers, developers

Around 2000, Steve Ballmer jumped across a Microsoft stage chanting one word, over and over: developers. People laughed. I spent years selling to developers, at Borland and at Microsoft, and I never laughed. Ballmer was right. Whoever wins the developers wins the platform.

That same year, at Stanford, a PhD student named Ian Buck was playing games. Seriously. He wired 32 GeForce cards together to run Quake in 8K. Then a question took hold of him. If these chips can do that much math for a game, why not for anything else?

Here is why the question made sense. Nvidia’s chips were built very differently from the processor in your laptop, the CPU. A CPU is a few brilliant generalists. Think of four star chefs who can cook any recipe, four plates at a time. A graphics chip, the GPU, is thousands of simple workers doing the same small task at once. A thousand line cooks who can only chop onions. Useless for a soufflé. Unbeatable for a mountain of onions.

A video game is a mountain of onions. A million pixels, redrawn sixty times a second, each needing the same small calculation. Underneath, that calculation is just math: multiplying large grids of numbers. Buck saw that science was full of the same math.

The problem was language. A chip only understands its own machine language, simple orders in 1s and 0s. Developers write in human-readable languages, and a translator, the compiler, converts their code for the chip. For graphics chips, the only languages available spoke “graphics.” Almost no scientist did.

So Buck built Brook, a way to write ordinary code for a graphics card. It became his thesis. Nvidia hired him in 2004 and paired him with its architect John Nickolls. Together they turned Brook into CUDA, released in early 2007. CUDA let any engineer who knew C, the most common language in engineering, put thousands of line cooks to work. Buck’s lesson from the field was simple: nobody wants to learn a new language.

Then came the go-to-market decision that changed everything. CUDA was free. And Nvidia put it on its chips across the line, gaming cards included. Every teenager’s gaming PC became a small supercomputer, and every student a future customer. From day one, Buck also ran a team building ready-made libraries, so a researcher never started from a blank page. His priority order: “Scientists first, programmers a very distant second”.

For years, the paying market was a small community of scientists. Who paid for the bet? Gamers. Huang called gaming the flywheel that funded huge R&D. Training followed, and universities started teaching CUDA. Nvidia’s Deep Learning Institute alone has trained hundreds of thousands of developers. Today, estimates of the CUDA community run from about 4 million to 6 million developers.

Ballmer chanted it. Nvidia built it, quietly, one scientist at a time.

Five years later, the thread pulls tight, in a bedroom in Toronto.

Before AlexNet, Krizhevsky had already written his own CUDA code to train neural networks on Nvidia cards. He was one of those students. Free software, a gaming card he could afford, libraries already there. Ian Buck’s community, five years after its birth, produced the moment that would change everything.

Training an AI model, it turns out, is the biggest mountain of onions ever seen. The same multiplication of large grids of numbers, repeated billions of times. Graphics was always math. AI is the same math, at a scale nobody had imagined.

Was it luck? Partly, and Huang says so. Had the team in Toronto been using AMD cards, he admits, the story might have gone differently.

But luck needs somewhere to land. In 2012, only one company had spent five years making its chips easy for scientists to use, and putting them in every student’s hands. In my reading, CUDA is the real reason the world was ready for AI. The bedroom was the spark. The community was the fuel.

In 2012, Nvidia was worth about $7 billion. Still, on paper, a gaming company.

Chapter 4: never stop at the current model

Most companies, handed a winning product, protect it. Nvidia kept climbing. Four stages, three leaps, each one selling a bigger problem solved, to a more senior buyer, at a higher price.

From chip to box. Huang dates the shift to deep learning to 2012. In 2016 it became a product: the DGX-1, a complete AI computer at about $300,000, pitched as the power of 250 servers in one box. A chip is a part an engineer chooses. A box is a solution an executive signs for.

In August 2016, Huang carried the first unit himself to a small nonprofit lab in San Francisco called OpenAI. He signed it in marker for the team. A small honesty note, in the spirit of this series: Huang now tells it as a launch nobody ordered. An April 2016 email to Elon Musk, filed in court, shows him writing at the time that demand was coming from everywhere. Founders polish their legends. Documents keep the dates.

From box to network. AI doesn’t run on one box. It runs on thousands of them talking to each other, and the connections between them become the bottleneck. In March 2019, Nvidia bought Mellanox, an Israeli networking company, for $6.9 billion in cash, outbidding Intel. It was almost half of Nvidia’s annual revenue and nearly all its cash. Huang said the plan out loud: from a chip company, to a server company, to a data-center-scale company. Mellanox made about $1 billion that year. In fiscal 2026, Nvidia’s networking business made more than $31 billion.

From network to AI factory. Then, in November 2022, ChatGPT. The world suddenly wanted what Nvidia had spent a decade assembling. Since then, its data center business has grown nearly 13 times. Nvidia now sells “AI factories”: entire racks of chips, networking and software that turn electricity into answers. Its newest platform, Vera Rubin, is in full production at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud and Nebius.

The gaming company is a memory. Data center is now 92.7% of sales. Revenue went from $60.9 billion in fiscal 2024 to $216 billion in fiscal 2026.

Chapter 5: many rivers to the sea

Success brought a danger every seller recognizes. A few very large accounts.

The big cloud providers, the Microsofts, Googles, Amazons and Metas of the world, represent about half of Nvidia’s revenue. In fiscal 2026, one direct customer alone accounted for 22% of revenue, another for 14%. And those same giants are designing their own AI chips. Your biggest clients slowly becoming your competitors: I have seen that movie in software, and it rarely ends well for the vendor who waits.

Nvidia didn’t wait. It opened new rivers, each with its own kind of buyer:

AI clouds. The “neoclouds” like CoreWeave and Nebius, which rent Nvidia computing by the hour to anyone who needs it. AI cloud revenue alone tripled year over year in early fiscal 2027.
Frontier AI partnerships. The labs building the most advanced models, OpenAI, Anthropic, xAI and others. Nvidia sells to them, works with them on future chips, and often invests in them.
Governments. Nvidia calls it sovereign AI: countries building their own national AI infrastructure. More than $30 billion in fiscal 2026, triple the year before, led by Canada, France, the Netherlands, Singapore and the UK. A whole new buyer persona: a nation.
Enterprise and industry. The long, patient business every platform eventually needs. Nvidia groups AI clouds, industry and enterprise together, and that group made $37 billion in a single quarter in early fiscal 2027.

Four rivers, four sales motions, four kinds of conversation. A cloud buyer counts cost per hour. A frontier lab wants the next chip before anyone else. A minister talks about national sovereignty. An enterprise wants proof it works for its business. Same company, very different pitches.

Chapter 6: lending your client the money to buy

Follow the money through this story and it turns completely around.

In 1993, Nvidia lived on $40,000 and $20 million of venture capital. In 1996, it lived on $5 million from a client who believed it. For the next two decades, gamers paid for its future.

Today Nvidia is the one writing the checks. In fiscal 2026, it invested $17.5 billion in private companies and infrastructure funds, including AI model makers that buy its products directly or through the clouds. In the first months of 2026 alone it committed more than $40 billion, including $30 billion in OpenAI and another $2 billion in CoreWeave. At the same time, it returned $41.1 billion to its own shareholders.

In plain sales terms: I help my clients afford my product, so they can grow, and buy more of my product.

Critics call it circular financing. Nvidia funds the buyer, the buyer buys Nvidia, and the revenue looks bigger than the real demand. Huang answers that these investments are a small share of what those companies must raise elsewhere.

I have built partner programs most of my career. Funding the ecosystem to speed up adoption is one of the oldest plays in the book. What is new is the scale, and the flexibility: equity in one place, help securing land and power for a cloud partner in another. At $40 billion, it is a growth engine. It also ties Nvidia’s health to the health of the clients it finances. Both are true.

Epilogue: a story of bets

Pull the whole thread and you find people. Human instinct: a CEO who chose the truth over the deal. Human innovation: a gamer at Stanford who wanted Quake in 8K, and a company that gave his idea to every student on Earth. Human intent: years of betting on developers before anyone paid for it.

What I keep from it:

The community was the real bet. Ian Buck’s CUDA, given away to every gamer and student, is what made a bedroom in Toronto possible. AI found Nvidia ready because Nvidia had spent years winning developers.
Solve the hardest equation with a partner. One letter to Hsinchu turned the impossible economics of chip factories into a thirty-year partnership.
Truth is a survival skill. A CEO flew to Japan to tell his biggest client the product was wrong. The client paid, invested, and tripled its money.
Never stop at the current model. Chip, box, network, AI factory. Each leap sold a bigger answer to a bigger buyer.
Open new rivers before the old one dries. AI clouds, frontier labs, governments, enterprises: four buyers, four pitches, one platform.
Be flexible with money. Help your clients afford you, so they can grow and buy more. Know the risk that comes with it.

Why I tell these stories

Markets tell this story as a stock chart. I read it as a story of human intent, human purpose and human innovation. That is what my work is about.

That is how I look at every company I work with, through three lenses. Commercial intelligence: what are you really selling, to whom, through which channels, and where is the money? A sense of human innovation: who are the people and ideas inside the company that can carry the next offer, the way Ian Buck carried CUDA? A sense of intent: what are you deciding to do, and when, while the market is still moving?

Most companies are not Nvidia. They face the same questions at their own scale, every quarter. That is the work I do, and this series is where I will keep pulling the threads behind the numbers.

If you want to pull the threads behind your own numbers, let’s talk. Book a conversation with me here: [YOUR CALENDLY LINK]

Sources

Palantir and the Misguided Hype Around AI Consulting

Em Richard — Sun, 13 Sep 2026 15:31:36 GMT

Palantir is often presented as one of the companies defining the AI era. Its vocabulary supports that perception: ontologies, artificial intelligence, operational platforms and sovereign infrastructure.

Behind this vocabulary, I see a business model that feels very familiar. The company sells a strong vision to senior decision-makers, places highly capable people inside the customer, integrates fragmented systems and adapts its technology to the organization’s specific needs.

This is consulting combined with proprietary technology.

The model has been executed particularly well. It may also have productized more of the consulting work than traditional firms have managed to do. My issue is with the hype that presents this as a completely new and inherently scalable technology model.

This matters beyond one company. Many businesses now describe customized services as platforms. Consulting becomes deployment. Customization becomes configuration. A representation of the customer’s business becomes an ontology. Consultants become forward-deployed engineers.

The terminology is different, but the business question remains the same. Can the company reproduce the value it creates without rebuilding the relationship, implementation and delivery process around every customer?

A good business creates a fair trade of value between two parties. A successful business makes that trade repeatable and scalable.

Palantir clearly has a good business. I am less convinced that it has made the trade as repeatable as the market assumes.

Consulting presented as technology

Technology consultancies have used a similar proposition for decades.

Your data is fragmented. Your processes are inefficient. Your existing vendors do not understand your business. There is value trapped inside the organization, and we can help you recover it.

Consultancies rarely sell technical features alone. They sell access to expertise and the confidence that they can understand the client’s business. Accenture does not simply promise to install software. It promises to help the client improve the business through a combination of technology, industry knowledge and organizational change.

The proposition here is structurally similar.

Other technology providers want to extract the customer’s data, capture its economics and make it dependent on their platforms. This company claims to do something different. It enters the organization, understands how it operates and builds a system around that reality.

The important positioning choice is to present the technology as the product and the consulting as a deployment method.

This is probably the smartest part of the model.

The business does not look like a consultancy supported by proprietary software. It looks like a software company whose product is so advanced that it requires embedded engineers to realize its potential.

Many technical founders are attracted to this approach. They believe they have built a powerful and flexible system, but customers struggle to understand it. They add services so the customer can experience everything the system is capable of doing.

The initial assumption is usually that the services are temporary. The company will learn from early implementations, improve the platform and gradually reduce the need for custom work.

In practice, the services often remain central. Customers value the system, but they also value the people interpreting their needs, adapting the technology and turning organizational ambiguity into technical decisions.

The technology becomes better while each significant customer still requires a new process of discovery, persuasion and implementation.

The forward-deployed engineering question

Forward-deployed engineers identify requirements, integrate data, map operational relationships, configure workflows and adapt the system as the customer’s needs become clearer.

From the customer’s perspective, this can look a lot like high-end technical consulting.

The company’s answer is that these engineers work from a common platform. They are not producing an isolated analysis or a custom application that begins from zero. What they learn during one deployment can become a reusable capability in the platform.

I think there is real value in this approach. It may offer a better system for capturing and reusing consulting work.

The problem is that a common platform does not automatically make the customer’s path to value repeatable. The technical components can be standardized while every deployment still depends on unusual access, institutional knowledge, personal judgment and organizational commitment.

Both sides can therefore be right. The company sees a reusable platform. The customer experiences a group of talented people making that platform work in its particular environment.

The economic question is whether each new deployment requires a proportional increase in highly skilled delivery work. If it does, the business is scaling largely through people. If the implementation effort declines while customer value remains high, then it is successfully converting experience into software.

A keynote or product demonstration cannot settle this question. We need concrete customer outcomes, implementation costs and evidence that similar value can be produced repeatedly.

This was what I found missing in Alex Karp’s recent AIPCon presentation. He spoke at length about value creation without providing a clear example of the value created. There was no detailed case showing the operational problem, what the platform changed, the measurable result and the amount of custom work required to achieve it.

The vision was clear. The proof was not.

Claiming the AI category

The business also positioned itself as an AI company before the public understanding of AI became dominated by frontier-model providers.

Before ChatGPT, Claude and Gemini, the term could describe a broad range of predictive systems, analytics, decision tools and operational optimization. Palantir occupied that space early.

Frontier models changed the category. If a general-purpose model can perform analysis, write code and interact with business data, customers may begin to question why they need another expensive platform between the model and their operations.

The strategic response is to turn the frontier models into the problem.

The models are powerful, according to the argument, but they are expensive to use because of token charges. They lack precision, industry context and knowledge of the customer’s operations. They introduce safety risks. They may also expose the proprietary knowledge that creates the customer’s advantage, which Karp calls its “alpha.”

The application layer and ontology are offered as the answer. A customer can use frontier models, open-weight models or fine-tuned models while keeping its permissions, workflows, decisions and institutional knowledge under its own control.

This is a credible technical proposition. It is also a traditional way of responding to a competitive threat.

When a new technology threatens to commoditize your advantage, you emphasize what it cannot do and the risks it introduces. You then position your own platform as the layer required to make the new technology useful.

The company does not need to build the best general-purpose model. It needs to stop the model provider from owning the customer relationship and the operational context surrounding it.

This dynamic is relevant to many technology-enabled service companies. AI makes their experts more productive, but it also makes some of their expertise more accessible to customers and competitors. The companies respond by moving their value proposition toward orchestration, governance, proprietary context and strategic judgment.

Sometimes this produces a defensible software platform. Sometimes it gives consulting work a new vocabulary.

AI may allow a company to perform customized work with fewer people and earn better margins. That is a valuable improvement. It does not necessarily make the work repeatable.

A relationship-led model

This growth model differs from the best-known software playbooks.

It is not product-led growth in the style of Lovable, where an individual can discover the product, experience value quickly and spread adoption through the organization.

It is not an open-source model like the one that helped Databricks develop its ecosystem around Apache Spark.

It is not primarily a displacement model like Snowflake’s, where the company can identify an existing category of infrastructure and offer a clear technical replacement.

This model begins with trust.

The customer must believe that the provider’s people can understand the organization, that its platform can represent how the organization works and that both can be trusted close to important operational decisions.

The customer has to provide access, explain internal processes, expose organizational problems and commit to changing how work gets done. The system is most likely to succeed when the customer is fully committed.

Palantir reported 1,049 customers as of June 30, 2026. Considering the potential size of its contracts, this is not necessarily a low number. It does tell us what kind of growth model the company has chosen.

The business does not need millions of independent users. It needs a relatively small number of powerful institutions willing to build deep and expanding relationships.

The unit of growth is the relationship.

This also makes its ideal customer difficult to define. The ideal customer is not simply a manufacturer, government agency or financial institution of a particular size. It is an organization whose leadership believes its operations are important and unique enough to justify an intensive technological partnership.

That quality is hard to capture in a conventional ideal customer profile. It is usually discovered and developed through relationships.

This model can produce exceptional customer outcomes. It can also become difficult to challenge.

Customers that succeed are treated as evidence of the platform’s power. Customers that do not succeed can be viewed as insufficiently committed, insufficiently sophisticated or unwilling to transform.

The company can become less accountable for making its value proposition simple and repeatable because customer belief is considered part of the implementation.

The model is attractive while the community is committed. It becomes fragile if the community loses confidence.

The founder’s role in maintaining belief

Karp is important to this model, but I do not think the company should be analyzed mainly as a profile of its founder.

His role is to create and maintain the belief surrounding the system.

His patriotism, attacks on conventional software vendors and repeated references to neurodivergence give the business a strong identity. Customers are invited to see it as more serious, independent and mission-driven than the rest of the technology industry.

I am still not sure what Karp means by neurodivergence in this context. He appears to use it as a broad description of unconventional, abrasive and intensely productive people. It becomes a way to present temperament as differentiation. Aggression is framed as intellectual independence. Refusing a customer request becomes integrity. Difficulty fitting into conventional institutions becomes evidence of unusual ability.

Patriotism provides an even stronger form of alignment. The company does not treat defense work as an uncomfortable source of revenue. It identifies with the mission of the United States and its allies.

This helps build trust in environments where institutional relationships, security experience and political alignment matter as much as technical performance.

Karp’s charisma supports the claim that this is more than a vendor. It is a partner that understands what is at stake.

Founder charisma often plays a similar role in relationship-led enterprise companies. The founder acts as the interpreter of the technology and the guarantor of its potential. Strong conviction bridges the distance between what the system delivers today and what it promises to deliver in the future.

This can be a real commercial advantage. It is also difficult to separate from the product’s success.

When trust depends on the founder, senior relationships and a committed community, these are part of the distribution model. They cannot be dismissed as branding around an independently scalable software engine.

What the numbers tell us

The financial results make it too simplistic to dismiss the company as a consultancy.

In the second quarter of 2026, it reported 93 percent year-over-year revenue growth, an 85 percent GAAP gross margin and a 47 percent GAAP operating margin. Its cost of revenue grew much more slowly than revenue. Customer count increased from 849 to 1,049, and no customer represented more than 10 percent of revenue.

These are exceptional economics for a company with significant implementation requirements.

They show that the business has created substantial operating leverage. They also suggest that the platform captures and reuses at least part of the work performed during customer engagements.

What the numbers do not show is whether the next thousand customers can be acquired and implemented through the same process. They do not tell us how much growth depends on specialized personnel, senior access, founder charisma or favorable political conditions.

High margins measure the ability to capture value. They do not by themselves prove that the value creation process is repeatable.

Michael Burry’s comparison of the company’s deferred-revenue-to-revenue ratio with Accenture’s is relevant here. The reported ratio of about 32 percent does not prove that Palantir is a consulting company. Deferred revenue excludes contracted amounts that have not yet been invoiced, and the business reported approximately $4.8 billion in remaining performance obligations as of June 30, 2026.

Government contracts and termination provisions also make direct comparisons with conventional subscription software difficult.

The ratio is still an interesting signal because it directs attention toward the commercial model. The company may have software margins without having a fully software-like method of acquiring, implementing and expanding customers.

The useful question is not whether it is secretly a consultancy. The distinction is unlikely to be that clean.

The useful question is how much of the value can be delivered repeatedly by the platform, and how much continues to depend on the flexibility of its people and relationships.

A good business with a strong system of defense

The company has built a strong system around its model. It has proprietary software, accumulated deployment experience, embedded customer workflows, high switching costs, institutional relationships and a committed community.

It also has a narrative that turns changes in the competitive environment into arguments for why the platform is necessary.

Frontier models become evidence that customers need its control layer. AI-related security concerns reinforce its experience with sensitive systems. Political concerns about sovereignty strengthen its patriotic positioning. The complexity of implementation supports the need for forward-deployed engineers.

Each potential threat is incorporated into the company’s defense.

This may sustain a very successful business for years. But a strong defense does not answer the question of scalability.

A business based on deep relationships and flexible delivery can become enormous. Accenture demonstrates that. Its growth simply operates under different constraints from a product-led software platform. It requires continued access, trust, skilled people and a willingness to adapt around the customer.

That flexibility is one of this model’s greatest strengths. It may also be one of its limits.

The lesson extends to every company now presenting consulting as an AI platform.

Proprietary technology can improve the quality and consistency of delivery. AI can make teams much more productive. Relationships can create customer loyalty and pricing power. A charismatic founder can sustain belief while the platform develops.

These are all real advantages.

They do not automatically make the value exchange repeatable.

To establish that, a company has to show that similar customers can increasingly reach similar outcomes with less bespoke intervention, less founder involvement and less dependence on exceptional commitment.

Otherwise, it has not eliminated consulting. It has built a better and more profitable consulting model.

Good is not the same as scalable

The value exchange is real.

Customers provide access, commitment and substantial economic value. The company provides technology, embedded expertise and the possibility of meaningful operational improvement.

For the right customer, this can be an excellent trade.

The misguided hype begins when that successful exchange is treated as proof of a universally scalable technology model.

The company may not have invented a completely new category of software. Its more meaningful achievement may be the construction of a consulting-led business capable of producing software-like margins while presenting the consulting as incidental.

That model is powerful, but it is also demanding.

It works when customers believe, participate and commit. It works when relationships provide access to the right institutions. It works when the people delivering the system can translate customer complexity into reusable platform capabilities.

The risk is not that the business creates no value.

The risk is that too much of the value still depends on the flexibility, intimacy and conviction of consulting, while the hype assumes the repeatability of software.

A good business creates a fair trade of value between two parties.

A successful business makes that trade repeatable and scalable.

Palantir has demonstrated the first. The second remains the question, both for the company and for every business now describing its consulting model as an AI platform.

This article analyzes Palantir’s business model and public financial disclosures. It is not investment advice.

Sources: Palantir Q2 2026 Form 10-Q; Palantir Q2 2026 business update; Palantir revenue and remaining-performance-obligation disclosure; summary of Michael Burry’s argument.