Forem: Karen Tam

$20 Billion Christmas Present: Why NVIDIA Acquired Groq to Crush the "Inference" Rebellion

Karen Tam — Sun, 04 Jan 2026 15:53:58 +0000

While the world was preparing for Christmas dinner, Jensen Huang was busy securing the next decade of NVIDIA’s dominance. On December 24, 2025, NVIDIA announced a shock move: a $20 billion deal to acquire the technology and core talent of Groq, the AI chip startup famous for its blistering inference speeds.

This isn't just another acquisition; it is a calculated strike to solve NVIDIA’s biggest existential threat. Here is the breakdown of why this deal happened and what it means for the AI landscape.

The "Ferrari for Uber Eats" Problem

To understand this deal, you have to look past NVIDIA’s current $4 trillion market cap. Jensen Huang knows that the AI market is splitting into two distinct eras:

Era 1 - Training (2020-2025): Building the models. NVIDIA dominates here with 86% market share.
Era 2 - Inference (2026-2030): Running the models (e.g., ChatGPT answering a user).

The Problem: Using NVIDIA’s H100/Blackwell GPUs for inference is overkill. It’s like "driving a Ferrari to deliver Uber Eats." It works, but it’s expensive and power-hungry.

Big Tech customers (Google, Amazon, Microsoft) know this. They have been aggressively building their own custom ASICs (Application-Specific Integrated Circuits) to cut costs. If NVIDIA ignores this shift, their market share is projected to collapse from 86% to 22.5% by 2030 as the world shifts from training to inference.

The Target: Who is Groq?

Groq isn't just another chip startup. It was founded by Jonathan Ross, the ex-Google engineer who invented the original TPU (Tensor Processing Unit)—the very chip that powered AlphaGo.

Groq’s "LPU" (Language Processing Unit) architecture is unique:

Speed: In benchmark tests, tasks that took GPUs 2 minutes were completed by Groq in 6 seconds.
Supply Chain Immunity: Unlike NVIDIA’s GPUs, Groq’s chips do not require CoWoS packaging or HBM (High Bandwidth Memory)—the two biggest bottlenecks in the global supply chain.

By acquiring Groq, NVIDIA isn't just buying a competitor; they are buying the "Father of the TPU" to weaponize his technology against his former employer, Google.

Deal Structure: The "Reverse Acqui-hire"

NVIDIA learned its lesson from the failed ARM acquisition, which was blocked by regulators in the US, UK, and EU. To bypass antitrust scrutiny this time, Jensen Huang used a clever playbook (similar to the Microsoft/Inflection AI deal):

No Full Acquisition: NVIDIA did not buy the Groq corporate entity.
Asset & Talent Transfer: NVIDIA paid $20B for non-exclusive licensing of the technology and hired the core leadership team (including CEO Jonathan Ross).
The Shell Remains: Groq continues to exist as a company with a new CEO, technically remaining an "independent competitor."

This allowed the deal to close in days, not years, leaving regulators with little ground to stand on.

The Fallout: Winners and Losers

This $20B check instantly reshuffles the silicon hierarchy:

Affected Party: AMD
AMD’s entire strategy with the MI300 series was to be the "value option" for inference. By absorbing Groq, NVIDIA now owns the world's fastest, most efficient inference technology. They have effectively removed AMD’s main wedge into the market.
Affected Party: Google
Google’s internal TPUs were their secret weapon. Now, NVIDIA owns the next evolution of that architecture. Jensen is essentially telling Google: "Anything you have, I have too—but mine is faster."
Complicated Winner: OpenAI
For Sam Altman, this is a double-edged sword.

Good: Inference costs will drop drastically using NVIDIA’s new LPU-based tech.

Bad: OpenAI is now even more trapped in the NVIDIA ecosystem. Jensen now owns the toll road for both Training and Inference.

The Paranoia of the Trillionaire

Jensen Huang often says:

"I wake up every morning feeling like we are 30 days from going out of business."

Most CEOs would relax with a 90% market share. Jensen saw a future where his customers became his competitors via custom chips, and he moved ruthlessly to stop it. He didn't buy Groq because he wanted to; he bought it because he refused to let anyone else have the "inference" crown.

The "Ferrari" now owns the "Scooter" fleet too. Good luck to the competition.

$2,300 Kill Switch: How One Hacker Saved a Headset and Broke a Tesla

Karen Tam — Sun, 04 Jan 2026 05:10:24 +0000

It was a ticking time bomb made of glass and silicon.

In 2018, early adopters paid $2,300 for the Magic Leap One, a futuristic augmented reality headset that promised to change the world. But by August 2023, the company had issued a death sentence: on December 31, 2024, the servers would go dark. Without them, a mandatory security check would fail, turning thousands of functional, high-end devices into expensive paperweights.

It was the ultimate example of planned obsolescence. Until a security researcher named Elise Amber Katze decided to intervene.

The Call to Action

The story starts in mid-2024, six months before the deadline. A friend approached Katze with a desperate plea: their expensive hardware was about to be bricked by a corporate kill switch.

For Katze, this wasn't just a technical challenge; it was a moral imperative.

"I morally dislike this," Katze told the audience at the 39th Chaos Communication Congress (39C3) in Hamburg. "Companies shouldn't turn functional devices into e-waste just because they want to sell newer devices."

With the clock ticking toward New Year's Eve, Katze went to work.

The Heist: Cracking the Silicon

The target was the NVIDIA Tegra X2, the system-on-chip (SoC) powering the headset. Katze didn't just want to bypass a login screen; she needed to own the hardware.

The attack unfolded in three cinematic stages:

The Backdoor (Sparsehax): Katze discovered that Magic Leap’s software was built on NVIDIA’s open-source code. She found a flaw in how the system unpacked files, allowing her to smash the stack and inject code over a simple USB connection.
The Foothold (Dtbhax): Getting in was one thing; staying in was another. She exploited the kernel’s loading process to ensure her jailbreak survived a reboot.
The "God Mode" (BootROM): This was the nuclear option. Using fault injection—literally glitching the hardware with voltage—she dumped the BootROM. She found a vulnerability in the chip's read-only memory. Because this code is etched into the physical silicon, it is unpatchable. NVIDIA cannot fix it with an update.

The kill switch was defeated. The headset was saved. But Katze wasn't done.

The Plot Twist

In a classic documentary twist, it turned out the Magic Leap One wasn't the only device relying on the Tegra X2.

While digging through the code, Katze realized the same unpatchable vulnerability existed in another piece of hardware—one that moves at 70 miles per hour.

Tesla Autopilot.

During her presentation, Katze dropped the bombshell: the exact same exploit chain she used to save an AR headset works on Tesla’s Autopilot 2 and 2.5 hardware. She demonstrated that an attacker with physical access could bypass the secure boot on a Tesla, granting them total control over the Autopilot module.

The Aftermath

What began as a mission to stop a $2,300 gadget from becoming e-waste ended with the exposure of a fundamental flaw in one of the automotive industry's most critical chips.

The Magic Leap servers shut down on December 31, 2024, as planned. But thanks to Katze, the devices didn't die. They were liberated.

"Netflix Effect": How Banks are Rewiring Money with AWS & NVIDIA

Karen Tam — Fri, 26 Dec 2025 10:07:55 +0000

At AWS re:Invent, the spotlight often falls on Generative AI for text (chatbots and summaries). But a quiet revolution is happening in the financial sector that is arguably much more valuable.

In a standout lightning talk, we learned how the world's financial giants—Stripe, Capital One, Nubank, and Visa—are moving beyond static rules. They are using the cloud to treat transaction data like a language, creating "Foundation Models for Money."

If you are interested in how Transformers and AWS infrastructure are modernizing the $1.4 trillion payments industry, here is your 5-minute summary.

The Problem: The "Static Rule" Trap 🕸️

For decades, banks have relied on static, "if-this-then-that" logic to detect fraud or approve credit (e.g., "If transaction > $5,000, flag it").

The problem? It’s rigid. It doesn't understand context. It treats you like a spreadsheet row, not a person with habits.

The new approach borrows from Netflix. Just as Netflix knows you'll hate a horror movie because you’ve watched three rom-coms this week, these new AI models learn your "spending personality" by reading your transaction history as a sequence.

The "Universal Remote" for Banking 🎮

The session highlighted a massive shift from "Siloed AI" to "Universal Foundation Models."

Instead of building one small model for fraud, another for marketing, and a third for credit limits, banks are training one massive "brain" on billions of transactions. This single model can then handle multiple downstream tasks.

Stripe: The Fraud Detection Leap 🛡️

Stripe recently unveiled their Payments Foundation Model. By moving from traditional models to this new architecture, the results were instant.

The Stat: They increased their detection rate for attacks on large businesses by 64% practically overnight.
The Tech: This model captures hundreds of subtle signals that specialized models miss, trained on tens of billions of transactions.

Nubank: Scaling on AWS ☁️

With 100M+ customers, Nubank is the perfect case study for cloud scale. They are training billion-parameter models using heterogeneous GPU clusters on AWS.

The Gain: They reported a 1.20% AUC lift (a metric for model accuracy).
Context: In the world of mature financial models, this is massive—roughly 3x the performance gain of a typical annual update.

Capital One: The Engagement Engine 🚀

Partnering with NVIDIA, Capital One showed that these models aren't just for defense; they are for growth.

The Result: By using transformer-powered recommendations, they saw a 35% improvement in predictions and a 10-12% increase in customer engagement.

Key Technical Takeaway: "Transaction Transformers" 🤖

The most exciting technical reveal came from Visa Research and their paper on "TransactionGPT."

They aren't just throwing text-based LLMs at numbers. They are designing novel architectures:

3D-Transformer Architecture: Visa designed a model specifically to handle the multi-modal nature of payments (Time + Amount + Merchant Type).
Sequence Modeling: The model hierarchically encodes individual transactions and their sequences over time, allowing it to predict future financial trajectories.

"We are moving from AI that recommends to AI that transacts." — Gurram Naveen, commenting on the shift toward Agentic Commerce

Why This Matters for Developers 🛠️

This is a "Self-Driving Car" moment for Fintech. The tech is proven by the pioneers, but it is not yet an off-the-shelf product. It requires:

Massive Compute: You cannot train these models on a laptop. You need the scale of Amazon EC2 P5 instances and NVIDIA H100s.
Custom Architecture: As Nubank noted, they had to build custom pipelines on top of their AI platform to handle sequence data.
No "Business in a Box": AWS and NVIDIA provide the tools (like NVIDIA NIMs and Amazon Bedrock), but the banks are building the proprietary "brains" themselves.

📚 Learn More & Get Started

Ready to dive deeper into Financial Services on AWS? Check out the source materials and tools:

Stripe Engineering: Read about how they apply machine learning to fraud detection. Visit Stripe Engineering Blog
Nubank's "Building Nubank": A deep technical dive into how they built foundation models into their platform. Read the Blog
Visa Research: Explore the work on deep learning and transaction security. Explore Visa Research
AWS for Financial Services: See how AWS is powering the future of banking and payments.Explore the Hub

Accelerating Cures: How IQVIA Scales AI with NVIDIA and AWS 🧬🚀

Karen Tam — Thu, 25 Dec 2025 17:26:38 +0000

At AWS re:Invent 2025, we often talk about "speed," but in the world of Life Sciences, speed literally saves lives.

In a fascinating interview on the expo floor, IQVIA (often called "the largest company you’ve never heard of") sat down with NVIDIA and AWS to discuss how they are dismantling the barriers to drug discovery.

If you are interested in how Agentic AI and GPU acceleration are reshaping healthcare, here is your 5-minute summary of the collaboration that is changing the game.

The Billion-Dollar Problem 💸

Lucas Glass, SVP of Technology at IQVIA, laid out the staggering reality of the pharmaceutical industry:

Time: It takes 10 to 15 years to bring a new drug to market.
Cost: The price tag sits between $1 billion and $2 billion.
Risk: The success rate is only about 10%.

The goal of this partnership is simple but ambitious: Shrink these numbers.

The "Triumvirate" of Innovation 🤝

The session highlighted a three-way partnership where each giant focuses on their "comparative advantage" to solve this problem.

1. IQVIA: The Domain Expert 🧠

IQVIA runs a massive portion of the world's clinical trials and acts as the largest healthcare data broker. They possess the data and the scientific know-how but realized they shouldn't be in the business of building data centers.

The Strategy: Use their massive historical data to design better clinical trial protocols (experimental designs) faster, predicting outcomes before a physical trial even begins.

2. NVIDIA: The Optimization Engine 🏎️

Lindy Wu from NVIDIA explained that their role goes beyond just selling hardware. Their goal is actually to lower the cost of compute.

The Tech: They provide the GPU-accelerated software stack.
The Key Innovation: NIMs (NVIDIA Inference Microservices). These are containerized, pre-optimized AI models that are "enterprise-ready." They allow IQVIA to take complex models and run them anywhere—specifically on AWS—without needing to manually optimize for the hardware.

3. AWS: The Scale & Infrastructure ☁️

Matt Carr from AWS highlighted the concept of "Undifferentiated Heavy Lifting."

The Role: AWS provides the global infrastructure and scale. IQVIA doesn't need to be an expert in HVAC, electrical engineering, or server maintenance.
The Integration: AWS provides the platform (like Amazon Bedrock or Amazon EC2) where NVIDIA’s NIMs and IQVIA’s data meet. This allows IQVIA to spin up massive compute power for a simulation and spin it down instantly to save costs.

Key Technical Takeaway: The Rise of "Agentic AI" 🤖

The buzzword of the session was Agentic AI.

We are moving past simple chatbots. In Life Sciences, "Agents" act as digital teammates.

Administrative Relief: A huge portion of healthcare costs is administrative (claims, denials, appeals). AI Agents can handle these repetitive workflows autonomously.
Clinical Design: Agents can analyze thousands of previous clinical trial protocols to suggest the optimal design for a new drug, turning a process that takes months into days.

"We don't want to take the human out of the loop... but how many administrators do you really want in a hospital system?" — Lucas Glass, IQVIA

Why This Matters for Developers 🛠️

Even if you aren't in healthcare, the architectural pattern here is a lesson for all builders:

Don't Reinvent the Wheel: IQVIA uses AWS so they don't have to build servers.
Optimize at the Edge: They use NVIDIA NIMs to ensure their code runs as efficiently as possible.
Containerize Everything: The ability to move models easily via containers (NIMs) allows for flexibility in deployment.

📚 Learn More & Get Started

Inspired to build your own Agentic workflows or explore Life Sciences on AWS? Check out these resources:

NVIDIA NIMs on AWS: Learn how to deploy NVIDIA Inference Microservices on Amazon SageMaker and Bedrock. [Read the Docs](#)
AWS for Life Sciences: Explore how AWS is powering the next generation of biology and clinical trials. [Explore the Hub](#)
Amazon Bedrock Agents: Start building your own autonomous agents today. [Get Started](#)

Did you catch the interview? What do you think about the shift from "Generative AI" to "Agentic AI"? Let us know in the comments!

My Takeaways from AWS re:Invent 2025: Bringing the Vegas Energy Home to Hong Kong 🇭🇰✨

Karen Tam — Thu, 25 Dec 2025 03:12:36 +0000

What happens in Vegas... definitely doesn't stay in Vegas when the AWS Hong Kong User Group is involved!

We just wrapped up our final meetup of the year, and for me, it was one of the most special sessions we’ve hosted. As I stood there looking at so many old friends and new faces, I realized that while re:Invent is a global phenomenon, its real impact happens when we bring those stories back to our local communities.

I had the privilege of sharing the stage with our UG leader Alex, AWS HK Hero Cyrus, Richard, and our AWS colleague Clifford. Together, we unpacked the massive announcements from AWS re:Invent 2025, but more importantly, we talked about the experience—the chaos, the learning, and the "Renaissance" of being a developer.

Here is my personal look back at our session and the key takeaways that stuck with me.

The "Fear of Missing Out" is Real (But That's Okay)

One thing that resonated with me during our panel discussion was the sheer scale of the event. We’re talking about 60,000+ builders.

I asked the guys, "Why go?" and "How do you survive?" Their answers were eye-opening.

The Hustle: Hearing Clifford talk about running a Game Day at 8:00 AM with over 300 participants—and seeing customers queuing up at 6:30 AM—reminded me of the intense passion in this community.
The Strategy: Cyrus and Richard shared a pro tip that I loved: Jet lag is your friend. Waking up at 5:00 AM in Vegas isn't a burden; it’s a superpower that gets you the best seats at the keynotes.

It made me realize that re:Invent isn't just about the sessions; it's about time management and making choices. As Clifford said, you will always have FOMO (Fear of Missing Out), but you just have to dive in and trust your schedule.

The Tech That Caught My Eye 👀

While I didn't play the "Builder Cards" tournament in Vegas this year (though we definitely played them in HK!), the technical updates our experts shared were game-changers. Here are the three announcements that I think will define our work in 2025:

The Era of the "Sleepless" Agent 🛡️

Richard introduced us to the AWS Security Agent. The concept is fascinating: an AI that acts like a "junior pen-tester" that never sleeps. It doesn't just scan; it reads code and understands logic. Hearing that it found vulnerabilities in a test app within two hours—things a standard scan missed—made me realize we are entering a new phase of agentic workflows where AI is an active teammate, not just a tool.

Simplifying the Complex with Durable Execution ⚡

I’ve seen how painful complex workflows can be for developers. Cyrus and Clifford discussed Lambda Durable Functions, and honestly, it sounds like a relief. The ability to write code-first workflows that can "checkpoint" and wait (for days, if needed!) without paying for idle compute is huge. It brings the focus back to the code logic rather than managing infrastructure glue.

Democratizing RAG with S3 Vectors 📂

We talk a lot about AI, but cost is always the blocker. The new S3 Tables with Vector Support is a massive win for accessibility. Moving vector search to a truly serverless, pay-as-you-go model on S3 means more builders can experiment with RAG (Retrieval-Augmented Generation) without spinning up expensive, dedicated databases.

The "Renaissance Developer"

We ended our session by watching the intro to Dr. Werner Vogels' keynote. It was emotional knowing this is his last keynote, but his message was powerful.

He spoke about the "Renaissance Developer." In a world where AI can write the syntax, our role is shifting. We need to be the ones who think in systems, who communicate, and who understand the "why" behind the "how."

Watching the video with the Hong Kong community, I felt a strong sense of solidarity. We aren't just coders; we are problem solvers. As Werner said, the tools change, but the builder spirit remains.

Looking Ahead

Hosting this recap wasn't just about reading release notes; it was about reliving the energy. From hearing about the "Replay" party (and that giant claw machine!) to discussing the future of AI, it was the perfect way to close out a year where we hosted 9 meetups and our very first Community Day.

I want to say a huge thank you to everyone who supported the AWS User Group this year. Whether you came for the pizza, the stickers, or the knowledge, you are what makes this community great.

Let’s grow even bigger in #2026! 🚀

Wishing you all a very Merry Christmas and a Happy New Year. See you in January!