Open-weight models are beginning to give application companies and enterprises something they have never had at this level: the ability to take powerful general intelligence, specialize it around their own work and reduce their dependence on the companies that built the AI frontier.
THE SIGNAL
For most of the generative-AI boom, the hierarchy has been relatively simple.
Companies such as OpenAI, Anthropic and Google build powerful general-purpose models. Software companies connect those models to specialized applications. Enterprises then pay to use the applications.
The intelligence largely originates at the top.
That structure is beginning to change.
Legal AI company Harvey has introduced Tenet, its first post-trained open-weight model. Harvey started with Kimi K3, developed by China’s Moonshot AI, and worked with Fireworks AI to adapt it for long-horizon legal work.
But the most important part of Harvey’s announcement was not the name of the model or where the foundation came from.
Harvey explicitly described one of its research goals as creating systems that allow law firms to build their own specialized models and “own their intelligence.”
That is the Deep Signal.
AI customers are beginning to acquire the technical ability to become AI builders themselves.
The emerging chain looks different:
Open-weight foundation model → specialized application company → proprietary workflows and expertise → industry-specific intelligence → enterprise-specific intelligence.
Every step potentially moves more control away from the original foundation-model provider and closer to the organization where the work actually happens.
WHAT THE MARKET IS MISSING
The AI industry is still frequently discussed as a contest over which company has the smartest model.
OpenAI versus Anthropic.
Google versus Meta.
American models versus Chinese models.
Benchmark against benchmark.
That competition matters, particularly at the frontier.
But it may not determine where all of the economic value eventually accumulates.
Harvey provides an early example of why.
In June, before Tenet was announced, Harvey reported that it had taken another open-weight model, GLM-5.1, and post-trained it specifically for legal-agent work.
On Harvey’s own Legal Agent Benchmark, the resulting system achieved a higher rubric pass rate than the versions of GPT-5.5 and Anthropic’s Opus 4.8 that Harvey tested. Harvey also found major improvements in grounding, specificity and tool use.
Those results come from Harvey’s own benchmark and should be viewed accordingly. They do not establish that the specialized model is universally superior to frontier models.
But they demonstrate something strategically important:
The strongest general-purpose model does not necessarily have to be the strongest model for every specialized job.
That changes the economics.
A company may not need to build a trillion-dollar frontier model if it can take an increasingly capable open-weight model and make it exceptionally good at the narrow set of tasks its customers actually pay it to perform.
The competitive question therefore begins shifting from:
Who owns the smartest general model?
to:
Who owns the best intelligence for the work being done?
Those are not the same thing.
FIRST-ORDER EFFECTS
The first effect is greater model choice.
Application companies such as Harvey do not necessarily have to send every task to the same outside AI provider. They can route difficult general reasoning to a frontier model while using specialized models for work where specialization produces better economics or performance.
That creates leverage.
If an application company can perform more work using models it controls or has customized, it potentially gains greater control over inference costs, latency, model behavior, deployment choices and product development.
The second effect is margin.
Every call to an outside proprietary model creates an upstream cost.
At small scale, that may be manageable.
At enterprise scale, with millions of documents, searches, agent actions and automated workflows, model usage can become a material expense.
Owning more of the specialized intelligence gives the application company another economic option.
The third effect is customization.
A general model must work across thousands of subjects.
A legal model can be optimized around contracts, litigation, due diligence, regulatory analysis, legal research and document-heavy workflows.
A financial model could optimize around financial statements, valuation, transactions and market data.
An insurance model could specialize around underwriting, claims and risk.
The closer the model gets to the actual work, the more specialization can matter.
SECOND-ORDER EFFECTS
The deeper change begins when the application company is no longer the final destination.
The enterprise itself may become the next model builder.
That is already beginning to appear.
The Financial Times recently reported that Latham & Watkins has purchased Nvidia GPU servers and is customizing open-weight models internally, giving the law firm greater control over client data, infrastructure and its choice of AI systems.
That takes the progression one step further.
Harvey specializes intelligence for law.
A law firm could eventually specialize that intelligence for itself.
Its documents.
Its transaction history.
Its preferred drafting structures.
Its precedents.
Its negotiating patterns.
Its internal procedures.
Its institutional knowledge.
Its risk tolerances.
Its accumulated expertise.
At that point, the organization is not merely adopting artificial intelligence.
It is beginning to encode part of the organization itself into artificial intelligence.
That may create an entirely new class of corporate asset: proprietary institutional intelligence that is not merely stored in databases but operationalized through models and agents.
And once organizations possess that capability, switching costs could change dramatically.
Today, a company may worry about being locked into one model provider.
Tomorrow, the more important question may be whether its proprietary intelligence can move between underlying models.
If the specialized layer can survive while the foundation underneath it changes, then the foundation model becomes increasingly interchangeable for some workloads.
That would fundamentally alter negotiating power across the AI stack.
WINNERS
Open-weight model developers gain relevance because their models become foundations other companies can modify rather than finished products that must compete directly with every closed frontier system.
Vertical AI companies gain leverage because they can combine customer relationships, domain expertise, proprietary workflows and increasingly their own specialized models.
Large enterprises gain optionality. Organizations capable of operating or customizing models internally may be able to choose between frontier APIs, specialized vendors and their own systems depending on security, performance and cost.
Infrastructure providers may also benefit. Training, post-training and operating specialized models requires chips, cloud infrastructure, inference systems, evaluation tools and deployment platforms.
And perhaps most importantly, organizations possessing valuable proprietary data and institutional knowledge may discover that those assets become even more valuable when they can be converted into specialized intelligence.
LOSERS
The pressure falls first on AI businesses whose only differentiation is access to someone else’s model.
If two applications call the same underlying model and neither possesses unique data, workflows, distribution or specialized intelligence, differentiation becomes difficult.
Frontier-model companies could also face a more complicated economic environment.
OpenAI, Anthropic and Google may continue producing the strongest general-purpose models in the world.
But they could increasingly find themselves serving as one layer inside a larger intelligence stack instead of controlling the entire stack.
That does not necessarily mean less demand for frontier models.
The opposite is possible: specialized systems may still call frontier models for their hardest tasks.
But the relationship changes.
Instead of:
“We need your model to run our product.”
the relationship can become:
“We will use your model when it is the best model for this particular job.”
That is a substantial difference in leverage.
WHAT HAPPENS NEXT
The most important signal to watch is whether Harvey’s experiment spreads beyond Harvey.
Watch for more vertical AI companies post-training open-weight models around specialized professions.
Watch for enterprises purchasing their own compute infrastructure or deploying models inside controlled environments.
Watch for software companies introducing model routers that automatically choose among proprietary, open-weight and internally trained models according to cost, security and performance.
And watch what happens to the data.
The next competitive battle may not simply concern access to larger training datasets.
It may concern access to the extraordinarily valuable information buried inside corporations: the decisions, documents, workflows and expertise accumulated over decades.
That information was created for humans.
AI gives organizations the possibility of turning it into operational intelligence.
There are serious limitations.
Training specialized models requires expertise, infrastructure, evaluation systems and clean data. Many enterprises will have neither the scale nor the desire to become AI laboratories.
Frontier models may also advance rapidly enough that renting intelligence remains cheaper and better for many tasks.
And open-weight models create their own security, governance and maintenance responsibilities.
So this is not a prediction that every company will build its own model.
The important change is simpler:
They increasingly have the option.
Options create leverage.
BOTTOM LINE
The first phase of generative AI concentrated power around the companies capable of building enormous foundation models.
The next phase may distribute some of that power back down the stack.
Harvey’s Tenet is one early example.
An application company took an open-weight foundation, specialized it around the work it understands, and began moving toward owning more of the intelligence behind its product.
Now major enterprises are experimenting with similar ideas.
That creates a progression worth watching:
Frontier intelligence becomes open-weight intelligence.
Open-weight intelligence becomes industry intelligence.
Industry intelligence becomes institutional intelligence.
And once that happens, the most consequential question in AI may no longer be:
Who built the smartest model?
It may become:
If every major industry eventually develops intelligence trained around its own data, workflows and expertise, does the greatest value remain with the company that builds the smartest general model — or move toward the company that owns the specialized intelligence where the work actually happens?
