• The GPU
  • Posts
  • Issue #5: ByteDance’s $7B Rumour, Meta's LCM, & TSMCs New Customers

Issue #5: ByteDance’s $7B Rumour, Meta's LCM, & TSMCs New Customers

Feat. ByteDance, Nvidia, Run:ai, Telekom Malaysia, Meta, TSMC, Broadcom, Digital Bridge, Microsoft, and Google

Author

Ben Baldieri
January 04, 2025 • Estimated Reading Time: 10 minutes

You’d think the year might start slow. It hasn’t.

ByteDance is allegedly planning to spend $7B on Nvidia GPUs. Meanwhile, Nvidia and Run:ai are open-sourcing parts of their marriage, Telekom Malaysia is trying its hand at GPUaaS, Meta is rethinking how AI processes language, and Atlanta is pushing back on data centre growth.

Every week, I break down what’s happening in GPU compute, AI infrastructure, and the data centres that keep it all running. If you’re looking to understand what matters and why, you’re in the right place.

Here’s what’s inside this issue:

  • ByteDance’s (denied) $7B bet on Nvidia Blackwell GPUs.

  • Nvidia finalising its $700M Run:ai deal and open-sourcing parts of it.

  • TM’s GPUaaS launch and what it says about AI demand.

  • Meta unveiling Large Concept Models to rethink how AI processes language.

  • TSMC’s silicon photonics leap with Nvidia and Broadcom.

  • A 240MW mega data centre in Atlanta, facing local pushback.

  • Microsoft’s VP of silicon engineering jumps to Google Cloud to lead chip development.

Let’s get to it.

The GPU Audio Companion Issue #5

Want the GPU breakdown without the reading? The Audio Companion does it for you—but only if you’re subscribed. Fix that here.

Top Story: ByteDance’s (Alleged) $7B Blackwell GPU Bet

ByteDance allegedly told suppliers it plans to spend $7B on Nvidia Blackwell GPUs.

According to the original story from The Information, suppliers were told the orders would likely be placed and deployed outside of China, thus circumventing U.S. export restrictions. ByteDance, naturally, denies the claims, calling the reports inaccurate.

Whether or not the deal is real, the buzz highlights the growing demand for GPUs in China despite U.S. restrictions and the strategic moves tech companies are making to secure computing power.

Why this matters:

  • Even Nvidia’s de-featured H20 GPUs are selling well in China, with demand reportedly increasing by 50% each quarter. ByteDance’s alleged $7B GPU purchase fits into this broader push for more compute power.

  • With the potential for a U.S. TikTok ban looming, ByteDance has every reason to double down on AI, leveraging the vast amounts of data it already has.

  • Nvidia is walking a tightrope: balancing massive demand from Chinese tech companies with U.S. export restrictions. If the rumours are true, expect Jensen Huang to add “Geopolitical Strategist of the Year 2025” to his résumé.

Nvidia & Run:ai Go Open Source

Nvidia has closed its $700M acquisition of Run:ai, a platform for GPU orchestration.

In a potentially unexpected twist, Nvidia and Run:ai are planning to open source aspects of the software and make it compatible with hardware from the likes of AMD, Intel, and others.

Why this matters:

  • Nvidia’s market dominance (90%+) can really only go in one direction.

  • By open-sourcing the software layer, they can still influence the overall direction of travel when that dominance invariably falls.

  • Nvidia is a software company, so with this move, we may yet see Nvidia AI software running on AMD hardware. All roads lead to Jensen.

Telekom Malaysia Pushes Into GPUaaS

Telekom Malaysia (TM) has launched a GPUaaS platform.

While it’s not exactly what you’d call a NeoCloud (it’s a massive telco, after all), the move shows how traditional industries can compete in the race for AI infrastructure. TM is also positioning this move as a way to “empower Malaysia’s AI aspirations”. South-East Asia is a hotbed of activity right now, with Siam.ai launching in Thailand and Nvidia planning to open a new R&D centre in Vietnam.

Expect more of this in the region in the future.

Why this matters:

  • TM’s move shows that the demand for AI infrastructure isn’t just limited to hyperscalers or startups.

  • The TM GPUaaS platform will be hosted entirely in-country as a foundation for building Malaysian sovereign AI capacity.

  • When telcos start offering GPUs, you know demand for compute isn’t just a niche anymore. It’s everywhere.

Meta’s Large Concept Models (LCM)

Meta’s FAIR team has introduced Large Concept Models (LCM).

LCMs are a new way to separate reasoning from language representation. By planning thoughts as humans do, LCMs aim to improve how AI processes and communicates complex ideas.

Why this matters:

  • LCMs focus on understanding concepts instead of just generating words, making AI better at handling complex tasks.

  • By mimicking how people think, Meta is pushing AI to be more adaptable and aware of context.

  • From customer service to research, this could reshape how AI handles more complex, language-heavy tasks.

TSMC’s Silicon Photonics Move

TSMC is pushing hard into silicon photonics.

With Nvidia and Broadcom as its first customers, silicon photonics are a big deal. The technology promises faster data transfer and lower power consumption, both critical for AI and high-performance computing.

Why this matters:

  • AI workloads need faster, more efficient data transfer, and silicon photonics delivers on that front.

  • Nvidia and Broadcom’s involvement shows this isn’t just theoretical - it’s being implemented now.

  • If viable at scale, silicon photonics could fundamentally change how AI infrastructure is built, potentially as early as this year.

Atlanta Balances Growth and Housing

Digital Bridge-backed Project Sail has begun work on a 1.5-million-square-foot data centre campus in Atlanta.

Why this matters:

  • AI workloads need huge facilities like Project Sail to scale, especially as demands increase.

  • As cities balance housing needs with tech growth, Atlanta’s pushback could signal a broader trend.

  • Despite Atlanta being a veritable DC hub, developers may start looking to other regions if restrictions tighten further.

Microsoft Exec Jumps to Google Cloud

Sundar is pinching Satya’s top people.

Google Cloud has hired Microsoft’s now ex-VP of Silicon Engineering to lead its chip technology and manufacturing. The move highlights Google’s increasing focus on custom silicon as it ramps up competition with AWS and Microsoft in AI and cloud infrastructure.

Why this matters:

  • Google’s move to strengthen its chip development team shows how central custom silicon is to hyperscaler strategies.

  • Losing a key executive to Google is a strategic blow for Microsoft at a time when competition in AI infrastructure is heating up.

  • Google’s focus on in-house chip development (and contracts with Broadcom for custom silicon) signals its intent to reduce reliance on a potential competitor like Nvidia.

The Rundown

Week one of 2025, and we’ve certainly hit the ground running.

ByteDance’s potential $7B GPU bet has everyone talking, while Nvidia’s Run:ai acquisition strengthens its position with an open-source twist. TM’s GPUaaS launch shows telcos are jumping into AI, and Meta’s Large Concept Models push language AI into new territory. TSMC is laying the groundwork for faster AI infrastructure, Project Sail is battling local resistance in Atlanta, and Google’s new chip exec hire shows the hyperscaler arms race is only getting more intense.

The year is already full of surprises. AI infrastructure isn’t slowing down. Neither am I.

Keep The GPU Sharp and Independent

Good analysis isn’t free. And bad analysis? That’s usually paid for.

I want The GPU to stay sharp, independent, and free from corporate fluff. That means digging deeper, asking harder questions, and breaking down the world of GPU compute without a filter.

If you’ve found value in The GPU, consider upgrading to a paid subscription or supporting it below:

☕ Buy Me A Coffee → https://buymeacoffee.com/bbaldieri

It helps keep this newsletter unfiltered and worth reading.

Reply

or to participate.