Enterprise AI Is Bottlenecked by Context, Not Model Capability

Ali Ghodsi argues that the central enterprise-AI problem is not a shortage of intelligence. It is that models do not possess the operational context that lets experienced employees make decisions, handle exceptions, and work through real organizational processes.

That argument is the practical consequence of Ghodsi’s more provocative claim that AGI already exists. He says today’s strongest models are already more capable than many people users encounter across a broad range of tasks, and that researchers he knew at UC Berkeley’s AMPLab in 2009 agreed current systems would meet the working definition of AGI held there at the time. In his view, the definition has continued to move as systems improve.

Yet he does not claim enterprises are already staffed by autonomous AI coworkers. He accepts the broad direction of reports suggesting that many enterprise proof-of-concepts fail, while questioning any precise headline figure. Inside companies, he says, people are still working through established processes with little visible productivity improvement.

The contradiction, as he sees it, is straightforward: models can reason, write, code, and answer questions, but they do not know how a particular organization works. Every company has people others consult because they hold accumulated context about its work. That knowledge may have built up over decades, but it is rarely available in a usable form to an agent.

For Ghodsi, this is why more spending on GPUs, data centers, or a future generation of frontier models does not automatically solve enterprise deployment. The work that matters is translating organizational knowledge, data, and processes into forms agents can use. With that context, he argues, models that already exist can address far more of the work than an isolated system answering prompts without knowledge of the company around it.

Apoorv Agrawal describes the challenge as having “the brain” while still needing to rebuild “the human body”: the hands, legs, and operational machinery through which a company gets work done. Ghodsi agrees. He calls the implementation problem human and organizational rather than a problem of model intelligence.

He situates AI in a familiar pattern of general-purpose technologies taking years to produce measurable productivity gains. Early PC users, in his telling, often used computers as typewriters: they typed documents, printed them, placed them in folders, and preserved the old workflow around the new tool. Factories initially did something similar with electrification, replacing steam engines with electric motors while retaining designs organized around line shafts. The larger gains came only after factories were redesigned around distributed electricity, larger floor plans, and independently operated equipment.

That is the analogy Ghodsi applies to AI adoption. Bank executives and other enterprise leaders tell him they know AI matters but cannot identify corresponding productivity gains in their organizations. Their processes remain configured for the old way of working.

The most concrete example came from Databricks’ process for building production connectors to systems such as Salesforce, Workday, or NetSuite. The conventional process took roughly three quarters—about nine months—to ship one secure, tested connector shaped by customer requirements. Ghodsi believed an LLM could write a connector in two days. The responsible team agreed that AI could help, but initially concluded it would reduce the cycle only from nine months to about seven and a half.

The breakthrough came when another employee re-examined the workflow from first principles. The team found that product managers spent a full quarter visiting customers and collecting detailed requirements before engineering began. It found that creating test environments for outside products was slow because Databricks was not specialized in setting up Salesforce, Workday, or NetSuite instances. It also found that each connector had a single owner, leaving projects vulnerable when that person was unavailable.

The redesign changed all three constraints. Rather than spending a quarter on exhaustive requirements gathering, the team compressed that work to one week and accepted that it could iterate later because software had become cheaper and faster to rewrite. It outsourced environment setup to firms that could provision those systems in parallel. And instead of assigning one person to one connector, it put seven people across seven connectors, reducing what Ghodsi calls a “bus factor one” dependency.

The reported result was seven connectors shipped in one quarter. The point was not that AI suddenly generated seven times as much code. The team altered its tolerance for incomplete requirements, its use of outside capacity, and its ownership model so that faster software creation could affect the whole delivery process.

Ghodsi’s conclusion is that a more capable future model would not, on its own, have produced the same outcome. Companies need to redesign how they collect information, provision systems, test software, allocate people, and make decisions. That is the practical route from model capability to productivity.

Ali Ghodsi rejects the proposition that AI makes software companies categorically obsolete. If software were dead, he argues, then OpenAI and Anthropic would be dead too: both depend on highly skilled people writing software. Nvidia would be dead as well, because its chip designs are also created by people using software. The more useful question is how AI changes the competitive conditions within software.

The valuation chart presented by Agrawal places the software-index median enterprise-value-to-revenue multiple at 4.1x, just below the 4.2x reference shown for the Covid period. The chart frames that level against prior drawdowns associated with Brexit, the “Tiger Tantrum,” and inflation and rates. Agrawal presents the current anxiety as a recurring market pattern: software has repeatedly been declared dead, then recovered. Ghodsi’s answer is that AI does represent a structural change, but not an extinction event.

First, the barrier to entry has declined because software is cheaper to create. New entrants can build products more quickly and at lower cost. But incumbents have the same new tool available to them; Databricks, he says, uses it as well.

Second, switching costs may fall as users increasingly interact through agents rather than through each product’s interface. A person who knows an iPhone, Gmail, Outlook, or Salesforce may resist changing because learning a new interface and moving data is costly. If an agent is carrying out the work across systems, Ghodsi argues, the underlying interface may matter less to the user.

That does not erase every moat. He points to economies of scale, brand, trust, security, certifications, patents, customer relationships, and proprietary data. A company with unique data still has an advantage even if a competitor can cheaply reproduce much of its visible software.

The more exposed companies are those that have coasted. Ghodsi says a company whose product has remained largely unchanged for a decade while revenue kept growing should be worried: a new entrant can now build something better much faster, and the incumbent may no longer have the internal capacity to innovate. Agrawal similarly identifies data-rich products embedded in core operational loops, including cybersecurity, as relatively resilient, while dated workflow products look more vulnerable.

Even those incumbents are not necessarily finished. Established companies have customers, data, and scale. If they build useful AI into their products and adjust pricing and cost structures where necessary, Ghodsi says, those advantages can remain substantial.

Apoorv Agrawal presents the current AI stack as an inverted version of cloud software economics: estimated AI semiconductor revenue is the largest layer, while application revenue is much smaller.

The figures describe the current distribution in Agrawal’s framework, not a forecast. The forecast belongs to Ali Ghodsi: he expects durable value to move to applications, particularly products that apply AI to expensive, consequential problems for which customers will pay.

He cautions that he is not giving investment advice and that identifying individual winners is difficult. If he were investing at the application layer, he says, he would make many early-stage bets and expect only a few to become major companies. The important point is not that every AI application will succeed, but that the most visible technical layer is not necessarily where lasting economics reside.

Ghodsi bases that view partly on technology history. Value, he says, repeatedly moves upward as underlying layers commoditize: from PC makers to operating systems, from software to virtualization, and beyond. His own experience in networking made him skeptical of assuming the hardest technical bottleneck becomes the largest business opportunity. Researchers once concentrated on multicast, or efficiently broadcasting a single live event to a large audience, when bandwidth was scarce. He started a company around the problem; then bandwidth prices fell and fiber expanded, undercutting the scarcity that made the technology seem urgent.

The consequential internet businesses instead emerged in applications that sounded mundane or strange at the time: selling books, hailing taxis, renting a bedroom, or sending short messages. AI’s major businesses, Ghodsi expects, may similarly be hard to identify while the market remains concentrated on chips, infrastructure, and frontier models.

Healthcare and education illustrate the kind of application he has in mind. Healthcare, he says, represents 17% of U.S. GDP and commands unusually high willingness to pay because people care intensely about their own health and that of their families. A company able to learn from a vast population of patients and use relevant genetic and clinical patterns to improve care could be extraordinarily valuable. In education, a category investors have often dismissed, an AI product that demonstrably improves learning could find demand because families care deeply about outcomes.

If such companies succeed, Ghodsi argues, their defensibility would not rest on code alone. They could accumulate proprietary data, economies of scale, and winner-take-most dynamics in particular markets. That is consistent with his enterprise argument: the durable advantage comes from combining a capable model with context, workflow ownership, and a problem important enough for customers to change how they work.

The model layer will remain valuable, he says, and centralized data centers will serve models at scale much as clouds do now. But open models will apply pricing pressure. He cites Moonshot’s Kimi 2.6 as an illustration of how quickly the frontier moves, saying it would have been the best model ever produced if released only months earlier, in January. His expectation is that frontier-model provision becomes an economies-of-scale business with thin margins, more like Amazon’s early book-selling business than a permanently protected monopoly.

Ali Ghodsi’s test for operators is not whether a model can complete an impressive task in isolation. It is whether the organization can expose the context and redesign the process that surrounds the task. Faster coding can produce only a marginal gain when requirements, test environments, handoffs, and staffing remain untouched.

For incumbents, lower software-production costs make inertia more dangerous without making existing advantages irrelevant. Customer relationships, data, trust, and scale can still matter; the question is whether a company uses them to build and operate a meaningfully better product.

For students and builders, Ghodsi’s advice is to avoid career decisions driven by fear of missing an imminent singularity or by the social-media consensus around the problem of the moment. He uses the history of multicast to illustrate how technical importance can be overtaken by changes in underlying economics. Conversely, he says, the internet existed for years before Airbnb began in 2009; foundational technology does not immediately reveal its most consequential applications.

His alternative is a long-term bet on a real behavioral or economic shift. He describes Jeff Bezos’s approach to the internet in those terms: starting with books, an unglamorous commodity, while betting that online purchasing would grow over time. The useful opportunity, in Ghodsi’s framing, may be less visible than the infrastructure or model race—but it must pair a durable demand with the willingness to refactor how the work is done.