- Towards AGI
- Posts
- Encrypted Data Fuels the AGI Race
Today, we’re diving into:
Gen AI: Banks Bet Billions On AI Strategy
Hot Tea: Encrypted AI: Compute Without Exposing Data
OpenAI: Alibaba Overtakes Google, Meta In AI
Closed AI: Meta's Open Model Splits The Industry
Big Banks Are Betting Billions On AI. Are You?
You've probably noticed AI headlines everywhere, but you may have missed how far ahead the biggest financial institutions already are. Several major banks are now spending well into the billions on AI every single year.
One bank alone has pointed to roughly two billion dollars in AI-driven savings, matching what it spent to get there. Another is raising its technology budget again in 2026, building on thirteen billion spent the year before.
Why This Should Worry You More Than Excite You
This isn't a pilot program anymore. It's a structural shift in how capital-intensive industries operate, and it's happening while many enterprises are still debating whether AI belongs in the core strategy.
Employees Are Already Ahead Of Leadership
Some institutions report that nearly ninety percent of their workforce is already using AI tools daily. Engineers report productivity gains of over thirty percent using generative tools on routine development work.
Adoption at the employee level is outpacing adoption at the strategic level, and that gap carries real risk.
The Gap That's Quietly Growing
When frontline teams move faster than leadership frameworks, governance becomes reactive instead of deliberate. That's how shadow AI usage and inconsistent risk controls creep into an organization unnoticed.
Governance Gaps Don't Wait
Every quarter without structured AI governance widens the risk gap your competitors are already closing.
What This Means For Your Strategy Table
The lesson for enterprise leaders isn't to copy what banks are doing. It's to treat AI as infrastructure, not initiative.
Build Your Own Playbook
Tie spend to measurable outcomes: Fund AI initiatives against specific cost or revenue targets, not general innovation goals.
Centralize ownership: Assign one accountable leader across business units instead of scattering AI decisions departmentally.
Close the adoption gap: Formalize training and governance before employee-level usage outruns leadership oversight.
Reassess workforce planning early: Model how AI changes hiring and role design before competitors force your hand.
Review vendor and partner strategy: Evaluate whether external AI partnerships accelerate you faster than building everything in-house.
The Real Risk Isn't AI. It's Standing Still.
Analysts are already questioning whether these massive technology budgets are producing enough return. That scrutiny will eventually reach every capital-intensive industry, not just finance.
The enterprises that treat this moment seriously now will set the pace their competitors are forced to follow later. The ones that wait will be reacting to a strategy someone else already wrote.
Your AI Can Now Work Blind, On Purpose
460+ open-source compiler tools built into one platform. Four peer-reviewed publications and counting. Years of research condensed into one open framework.
A new open-source toolchain now lets AI models process encrypted data without ever seeing it in plain form. It converts standard AI models into ones that compute directly on encrypted inputs.
The project has been quietly maturing since 2023, and it just crossed a threshold worth every leader's attention.
Why This Isn't Just A Developer Story
Every enterprise handling sensitive data has faced the same tradeoff: use AI and expose data, or protect data and limit AI. This closes that gap.
How The Technology Actually Works
Homomorphic encryption allows computation on data while it stays encrypted the entire time. The output is only meaningful once decrypted by the authorized party.
Google's engineering team built the framework specifically to make this practical, not just theoretical.
The Cost Barrier Is Finally Dropping
This approach has historically been too slow for real use. Google's own team acknowledges the technique still carries a "nontrivial cost overhead," according to its engineering blog post announcing the project.
Even so, Google frames it as a cost problem rather than a capability problem, and says that cost is falling fast enough to make it viable for healthcare and finance workloads.
Where This Is Already Being Tested
Early applications include fraud detection, network intrusion detection, and private recommendation systems, all running on data that stays encrypted throughout processing.
Developers can write standard code and compile it into encrypted-data versions, without needing deep cryptography expertise themselves.
How Enterprises Can Turn This Into Growth
This isn't just a security upgrade. It's a new way to unlock data you've been sitting on out of caution.
Revisit data you've avoided using: Encrypted-computation tools mean regulated or sensitive datasets can now power AI safely.
Reduce compliance friction: Processing data without exposing it can simplify audits across healthcare, finance, and cross-border operations.
Pilot before scaling: Start with one high-value, high-sensitivity use case rather than a full infrastructure overhaul.
Upskill your technical teams: Familiarity with privacy-preserving computation will become a real competitive skill, not a niche one.
Watch vendor ecosystems form: Open frameworks like this one usually spawn commercial tools fast. Early evaluation gives you first-mover pricing leverage.
Your Data Is Ready. Is Your Strategy?
Turning locked-away data into usable, governed AI assets takes more than encryption. It takes a strategy built for it.
The Advantage Goes To Whoever Moves First
Privacy-preserving AI has been a promise for years. This shift makes it a practical option enterprises can actually pilot today.
The leaders who explore it early will unlock data their competitors are still too cautious to touch.
The AI Race Just Changed Leaders
3 billion+ model downloads in six months. 460+ open-source models. 300,000+ derivative models built on top.
Alibaba's open-weight AI models have quietly become the most downloaded in the world, ahead of both Google and Meta.
Google reported 418 million downloads for the year. Meta reported 227 million. Alibaba passed both in a fraction of the time.
Why This Number Actually Matters
Downloads signal something deeper than popularity. They show which models developers trust enough to build products on top of.
Open Source Is Becoming The Default Choice
Alibaba's Qwen models are increasingly treated as a starting point rather than an alternative. Developers are fine-tuning and deploying them as standard workflow.
That shift is exactly why American firms are responding fast.
Meta And Nvidia Are Already Reacting
Both companies recently released new open models of their own, signaling that the open-versus-closed debate is now a competitive front, not a philosophical one.
The Real Question For Your Boardroom
This isn't really about which company wins the download count. It's about whether your enterprise still controls its AI infrastructure decisions.
Open models offer lower cost and more flexibility, but they also shift more infrastructure responsibility onto the enterprise adopting them.
How Enterprises Can Get Ahead Of This
Leaders don't need to pick a side today. They need a framework for evaluating both paths clearly.
Audit your current AI stack: Identify where you depend on closed models versus where open alternatives could reduce cost.
Separate hype from utility: Evaluate models on performance for your specific workflows, not headline download counts.
Build internal AI governance early: Open adoption increases your responsibility for security, so formalize oversight before scaling usage.
Diversify vendor dependency: Avoid anchoring your entire roadmap to a single provider, open or closed.
Watch procurement and compliance exposure: As AI moves into finance and treasury functions, infrastructure choices carry real operational risk.
The Shift Is Already Underway
Whether your enterprise leans open or closed, the competitive map is being redrawn in real time.
The businesses that treat this as an infrastructure decision, not a trend, will be the ones setting terms instead of following them.
Silicon Valley Just Split In Two
One month between flagship release and open-source release. One consumer-grade GPU needed to run it. Four major tech companies now on the open side of this divide.
A major AI lab just released an open-source version of its most advanced model, and it runs locally on a single consumer GPU.
No distant data center required. That detail alone signals a real shift in how AI power could be distributed going forward.
Why Everyone In Tech Is Watching This Fight
Meta has pushed this open-source release as part of a much larger industry split, one that could reshape who controls AI going forward.
The Divide Behind This Release
Some major labs believe AI should stay closed, developed, and controlled by a small number of companies for safety reasons.
Others, including Meta, Nvidia, Google, and Microsoft, argue that closed AI concentrates too much power in too few hands.
Cheap Chinese Models Forced This Decision
Inexpensive open-source models out of China have gotten close enough to top-tier performance that competitors could no longer ignore the pressure to respond.
What Makes This Release Different
The new model is functionally identical to Meta's flagship version, capable of generating code, text, and images without relying on cloud infrastructure.
That local-first design is what turns this into a genuine strategic option for enterprises, not just a research release.
A Return, Not A First Attempt
Meta had backed away from open-source AI when rivals pulled ahead. This release marks a deliberate return to that earlier strategy.
How Enterprises Can Turn This Into Growth
Locally runnable models change the cost and control equation for any enterprise handling sensitive workloads.
Evaluate on-device deployment: Running models locally can cut cloud dependency and reduce recurring infrastructure costs.
Reassess data residency risk: Local execution keeps sensitive data inside your own environment instead of external servers.
Test before committing: Pilot the model on a contained workload to measure real performance against your current stack.
Track the open versus closed divide: Vendor strategy here will shape pricing and licensing terms industry-wide for years.
Build internal capability now: Early familiarity with open models gives your technical teams leverage in future vendor negotiations.
The Advantage Goes To Whoever Adapts First
This isn't just a research milestone. It's a live test of whether open models can genuinely compete at the top tier.
Enterprises that explore this shift early will have more control over cost, data, and infrastructure than those who wait for the debate to settle.
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