- Towards AGI
- Posts
- Inside The $6B Move Toward Open AGI
Today, we’re diving into:
AI news: Enterprise AI Agents Reach Deployment Stage
Hot Tea: AI Model Licensing Carries Legal Risk
OpenAI: Cut AI Costs Without Losing Quality
Closed AI: Nvidia's $6B Bet Reshapes AI Sourcing
Your AI Agent Rollout Is About To Get Real
A new industry study found that 98% of employees inside one leading AI lab use agentic tools every day, yet fewer than 20% of that lab's business customers have adopted the same product. Individual subscribers barely cross 1%.
That gap is not a rounding error. It is the clearest signal yet that agentic AI inside your organization will not spread the way the demos promised.
Fix Your Data Blind Spot
Every agent rollout exposes the same weak link: ungoverned data. See what's hiding in yours before a competitor does.
The Agent Gold Rush Nobody Is Winning Yet
OpenAI just launched its enterprise agent product, designed to give non-technical teams the same autonomous power that software engineers already have with coding tools.
The pitch sounds simple. Give a model access to your inbox, your files, and your workflow tools, and let it complete entire projects without supervision.
Vertical-specific rivals in law, sales, and finance are already chasing the same customers with a model-agnostic approach, plugging in whichever engine performs best at any given moment.
Why Your Teams Aren't Adopting Fast
The obstacle is not intelligence. It is trust. Engineers who built the tool are comfortable handing over calendars, chats, and financial data. Most leaders are not.
Permission setup is often confusing, and partial access rarely works. Full access is usually the only path to functionality, which is a hard sell for compliance teams.
The Real Bottleneck Is Not The Model
Every agent runs on what engineers call a harness, the layer of software that decides what an AI can see, which tools it can touch, and how much it explains itself.
A minimal harness assumes the model is smart enough to work alone. A guided harness checks in with the user at each step. Most enterprise buyers still prefer the second.
What Harness Design Means For You
Unlike a plug-and-play SaaS tool, agent harnesses shape risk directly. Fewer checkpoints mean faster output but less oversight, and that tradeoff belongs on your leadership agenda, not IT's.
Vendors rarely explain this distinction upfront. It shows up later, in audit reviews, security incidents, or a quiet loss of trust once an agent acts outside its intended scope.
How Enterprise Leaders Can Get Ahead
Waiting for the tooling to mature is a losing strategy. The organizations pulling ahead are treating agent adoption as a governance project, not a software rollout.
Pilot with checkpoints: Start agents on low-risk, high-volume tasks like reporting or scheduling before touching sensitive data.
Own the permission model: Assign a single team to define exactly what each agent can access, rather than leaving defaults to vendors.
Track token spend early: Usage costs can outpace subscription fees fast, so build cost visibility before scaling company-wide.
Choose harness over hype: Evaluate how much oversight a tool builds in, not just how capable the underlying model claims to be.
The lesson from this current wave of agent products is simple. The model matters less than the guardrails wrapped around it, and the enterprises that design those guardrails first will be the ones actually capturing the value everyone else is still chasing.
Your AI Model License Might Be A Trap
A major open-source foundation just asked regulators to formally certify a new AI licensing standard, one already adopted across four model families spanning agentic AI, robotics, and simulation.
The license was co-built with four of the biggest names in enterprise technology. Yet weeks after submission, reviewers still cannot agree on what it actually obligates you to do.
The License Nobody Can Fully Agree On
The Linux Foundation submitted a license called OpenMDW, covering models, data, and weights, for official open-source certification this month.
The license itself is not new. It launched months earlier, built alongside Amazon, Meta, IBM, and Microsoft, and one major chipmaker already applies it across its flagship model lines.
That adoption already spans agentic AI, quantum computing, robotics, and simulation products, meaning the license is shaping deployments well before regulators finish reviewing it.
Why This Matters For Your Stack
If your teams already build on open models carrying this license, you are operating under terms the industry itself has not finished defining.
We want system administrators to start integrating them into their systems immediately.
Why Old Licensing Rules Don't Fit New AI
Standard open-source licenses were written for readable code. AI models are mostly numerical weight files, produced by training runs that often draw on data nobody fully discloses.
That mismatch is not cosmetic. Reviewers are still split on whether redistributing a model's weights carries the same legal weight as redistributing its source code.
The same debate played out once already when the Open Source Initiative wrote its AI definition, drawing criticism from companies on both sides of the openness line.
The Open Question Sitting On Your Desk
A second unresolved question matters even more for procurement teams: whether calling a model open also requires the vendor to release its training data.
Until that question is settled, any contract built on the assumption of full openness carries a gap that could surface later as a compliance dispute.
How Enterprise Leaders Can Get Ahead Of This
Waiting for the Open Source Initiative to settle this debate leaves your legal exposure entirely in someone else's hands. Leading enterprises are acting before certification lands.
Procurement teams that assume "open source" means fully risk-free are the ones most likely to face disputes once formal certification rulings finally arrive.
Audit your model licenses now: Map every AI model in production against its exact license terms, not just its "open" label.
Separate weights from data risk: Treat trained weights and training data as distinct legal categories when reviewing vendor contracts.
Loop in legal early: Bring counsel into AI procurement conversations before deployment, not after a dispute surfaces.
Watch the certification outcome: Assign someone to track the review, since the final ruling will reshape vendor obligations industry-wide.
Licensing ambiguity rarely stays theoretical for long. The enterprises that map their exposure today, rather than after certification closes, will be the ones setting terms instead of scrambling to meet them.
A clear internal policy on model licensing also protects you from a second risk: vendors quietly rewriting their terms once certification decisions eventually land.
Your AI Bill Is About To Get A Reality Check
One major telecom just cut its AI spending on coding and other advanced tasks by as much as 56%, and the quality of the output dropped by only 2%.
The company now routes 40% of employee queries to open-source models, and it plans to push that share to between 60% and 70% within a few years.
The AI Bill Nobody Wants To Explain Twice
AT&T just proved that most enterprise AI spending is not tied to output quality. It is tied to habit, and habit is expensive.
The company built model routers that judge how complex a task actually is, then send simpler requests to cheaper models instead of the priciest frontier option by default.
That single shift protected performance while trimming costs on coding and other advanced workloads, proving the two goals are not actually in conflict.
Why This Should Worry Every CEO
If your teams are still sending every query to the most expensive model available, you are likely paying frontier prices for tasks that never needed frontier power.
That gap between actual need and default spend is exactly where AT&T found its savings, and it is almost certainly hiding inside your own AI budget too.
Open Source Is Closing The Gap Fast
AT&T is leaning on open-weight models from Nvidia, Meta, and Google to cut its reliance on premium subscriptions from the leading AI labs.
The company has historically seen open-source models lag frontier releases by six to ten months, but that gap keeps shrinking with each generation.
Notably, the company is holding off on open models from Chinese developers while it evaluates the risks, a caution worth copying in your own vendor review.
Just as good or better" than older frontier models.
How Enterprise Leaders Can Cut Costs Without Cutting Quality
The shift from flat subscriptions to token-based billing means unmanaged AI usage now hits your budget every single month, not just at renewal time.
Analysts have already flagged that the era of pushing every employee toward the biggest model for every task is coming to an end industry-wide.
Deploy a model router: Match task complexity to model cost instead of defaulting every query to the most expensive option.
Pilot open-weight models: Test leading open-source options on lower-risk workloads before shifting mission-critical tasks.
Set a usage ceiling: Cap per-employee token spend and review it monthly, not annually.
Vet every model source: Evaluate security and data risk before adopting any open-weight model, regardless of origin.
Track cost per query: Measure spend at the task level so leadership sees exactly where budget is going, not just the total bill.
Cut Your Data Waste
Uncontrolled AI spend often starts with uncontrolled data. See where duplicate and unmanaged records are quietly driving costs.
The era of unchecked AI spending is ending. Enterprises that build smarter routing and open-source strategies now will control their costs while competitors keep overpaying for power they rarely use.
The lesson is not to abandon frontier models entirely, but to reserve them for the work that genuinely demands that level of capability and cost.
Your AI Model Choice Just Got A $6 Billion Twist
One chip giant just committed $6 billion to license AI technology from a single startup, plus another $1 billion in direct investment at a $12 billion valuation.
The goal is a U.S. open-weight model built to rival Chinese systems that have already logged more than 3 billion downloads in six months.
That chip giant is Nvidia, and the bet signals a much bigger shift in how enterprise AI budgets will be allocated over the next few years.
The Chip Giant Betting Big On Open Models
One leading chipmaker is licensing technology from AI startup Poolside and absorbing more than 100 of its engineers into its own open-weight model project.
The move pushes that company beyond hardware and into the model layer itself, a shift that changes what buying its chips will mean for your enterprise stack.
Nvidia, the company behind this deal, has spent years supplying the infrastructure underneath the AI race. Now it wants a stake in the models running on top of it.
Why This Reshuffles Your Vendor Map
If you procure AI infrastructure and models separately today, that separation is starting to blur, and your contracts should reflect that shift soon.
China's Open Models Forced This Move
Chinese developers have built a commanding lead in open-weight AI, with downloadable models that companies can customize and run on their own servers.
Several major U.S. firms have already adopted Chinese open models hosted locally, citing aggressive pricing that closed, subscription-based options struggle to match.
That competitive pressure is exactly what pushed a hardware supplier to fund its own frontier-grade open model instead of waiting for someone else to build one.
Open weights and American AI leadership will be judged by whether the U.S. builds an ecosystem that spreads into every sector.
How Enterprise Leaders Can Get Ahead Of This Shift
Model sourcing is no longer a back-office decision. It now touches cost, national exposure, and long-term vendor lock-in, which puts it squarely on your agenda.
Waiting to see how this plays out is itself a decision, one that leaves your procurement team reacting instead of setting the terms.
Map your model origins: Know exactly which models power your workflows and where their weights and training data originated.
Watch the open-weight race: Track new open-weight releases alongside Chinese alternatives before locking into long contracts.
Diversify deliberately: Avoid depending on a single vendor for both infrastructure and models, since that combination now concentrates risk.
Revisit procurement policy: Update vendor criteria to weigh openness, cost, and geopolitical exposure, not just raw performance benchmarks.
This funding fight is really about who controls the default choice for enterprise AI. Leaders who track it now will not be surprised by what comes next.
The companies that treat model sourcing as a strategic decision, rather than an IT detail, will end up dictating terms instead of just accepting them.
Journey Towards AGI
Research and advisory firm guiding on the journey to Artificial General Intelligence
Know Your Inference Maximising GenAI impact on performance and Efficiency. | Model Context Protocol Connect with us, and get end-to-end guidance on AI implementation. |
Your opinion matters!
Hope you loved reading our piece of newsletter as much as we had fun writing it.
Share your experience and feedback with us below ‘cause we take your critique very critically.
How's your experience? |
Thank you for reading
-Shen & Towards AGI team