Databricks CEO Ali Ghodsi: Most Companies Don't Need Stronger Models
Databricks co-founder and CEO Ali Ghodsi stated that if cutting-edge models were to stop improving, the vast majority of companies would not even notice, as the models are already smart enough; what they lack is internal organizational context, not higher intelligence.
He pointed out that models do not attend every meeting, do not know what everyone is thinking, and do not grasp all processes. In every organization, there are always a few employees who understand the big picture, and when they leave, it causes anxiety among everyone; models lack this context. If this context were injected into existing models, organizations worldwide could gain significant productivity without needing models that can better solve Navier-Stokes equations, conjecture, or score higher on humanity's last exam.
Ghodsi summarized the issue: AI does not have an intelligence problem, but a context problem. In businesses, agents fail to get things done because the questions and data are incomplete; if proprietary data, permissions, and processes are fed into the models, they can provide correct answers. Most organizations are still stuck at chatbots and coding assistants and have not yet moved to using agents for bulk automation; the adoption curve itself lags behind model capabilities.
This assessment aligns with their product line: the company uses lakehouse architecture to store structured and unstructured data, then connects business terminology, entities, and metrics into a model-callable enterprise ontology using genie ontology, interfacing with OpenAI models under controlled permissions. Databricks itself was previously reported to have a valuation of about $190 billion and emphasized that it is not in a hurry to go public.
He also acknowledged that the company spent about $20 million on cluster expenses while training the open-source model dbrx, with only about $4 million being effectively used for training, while the rest was wasted on trial and error, making them even less willing to allocate resources to compete on intelligence benchmarks with cutting-edge labs.
In market mechanisms, this redirects capital from "the next version of a larger model" to "who can integrate meetings, emails, and table permissions into the model." What is sold is a narrative of pure inference token growth, while what is bought is data platforms, governance, and ontology layers. Funding flows towards lakehouses, permission audits, and agent orchestration; beneficiaries are infrastructure companies that hold access to enterprise data, while those under pressure are model labs that rely solely on benchmark scores and cannot enter client internal systems.
Corporate procurement is still constrained by privacy, auditing, and competitive filtering, and the speed of context injection depends on whether clients build their own ontologies, rather than just buying a software package to complete the task.
Source: Public Information
ABAB AI Insight
Ghodsi redefined AGI as "smarter than those around most of the time," and with this definition announced that AGI has arrived, not for philosophical debate, but to shift budgets from training clusters to lakehouses. After Databricks acquired Mosaic and promoted dbrx, it publicly acknowledged training waste, effectively using its own bills to prove that merely burning cards to pile up benchmarks does not solve the issue of clients lacking agent colleagues.
Capital is being directed to three areas: proprietary data in lakehouses, ontology layers, and distribution channels that confine GPT-like models within permissions and audits. Clients are not looking to solve conjectures; they want to replicate "that all-knowing old employee" into a callable object. Whoever controls meeting notes, emails, and table-level permissions can enhance the marginal output of existing models without waiting for the next scorecard.
A similar path is taken by Snowflake and Palantir, which use governance layers to block model vendors from directly connecting to client data, and ServiceNow, which writes ticket processes into executable agent objects. The industry is in a control phase: model intelligence is excessive, organizational adoption is insufficient, and the bottleneck has shifted from parameter quantity to context ownership.
This represents an interface shift in technological substitution. The mechanism is that intelligence has been commoditized, while context remains organizational property; benchmark exams can be publicly gamed, but meeting minutes cannot. When productivity depends on who can legally read internal context, pricing power shifts from model labs to data platforms.
ABAB News · Cognitive Laws
- Once models are smart enough, what they lack is presence.
- Key employees in organizations are models that have not yet been written into data.
- Rising benchmark scores cannot replace meeting minutes.