Issue #13: DeepSeek, HPE & Juniper vs. The DOJ, & 20GW of Natural Gas

There’s nothing quite like waking up on Monday to a trillion-dollar stock market rout.

The infrastructure race continues, but this week, we got a stark reminder that efficiency breakthroughs can upend entire markets. DeepSeek’s latest model sent shockwaves through AI stocks, Meta pledged ‘hundreds of billions’ to AI, and Lambda is securing Nvidia GB200 racks. Meanwhile, the US is scrambling to power 20GW worth of data centres, ASML is proving chip demand isn’t slowing, and the DOJ is stepping in to block a $14B networking megadeal.

Never a dull moment.

Every week, I break down the moves shaping GPU compute, AI infrastructure, and the data centres that power it all.

Here’s what’s inside this week:

• DeepSeek’s cost-efficient AI model disrupts the industry.

• DOJ moves to block HPE’s $14B Juniper acquisition.

• Lambda’s Pegatron partnership for Nvidia GB200.

• Alibaba vs DeepSeek: China’s AI competition heats up.

• ASML’s chip sales surge despite AI market volatility.

• 20GW of natural gas planned to power AI data centres.

• Meta commits ‘hundreds of billions’ to long-term AI infrastructure.

Let’s dive in.

The GPU Audio Companion Issue #13

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

DeepSeek’s Cost-Efficient AI Model Disrupts the Industry

DeepSeek’s R1 model sparked chaos in AI markets, raising questions about cost efficiency and the future of AI infrastructure.

Chinese AI start-up DeepSeek made waves after unveiling an AI model it claims was developed for just $6 million, a fraction of what companies like OpenAI and Google spend on model training. Though that claim is highly dubious, the announcement triggered the single largest one-day loss in Nvidia’s history, wiping $589 billion off its market cap, before recovering $260 billion the next day. The entire market is now questioning whether AI’s future will still be driven by massive GPU-intensive training or whether DeepSeek’s cost-cutting breakthroughs signal a paradigm shift. Sceptics argue the true cost could be much higher, factoring in prior research, but the impact was clear:

AI efficiency is now the battleground.

Why this matters:

• DeepSeek’s model suggests AI training costs could be far lower than previously assumed, though SemiAnalysis have already shown training costs for R1 were likely far higher than reported.

• Even if the training cost claims are false, the efficiency gains and true open source release of the reasoning capability have massively lowered barriers to entry into the SOTA model market (and arguably stolen control of open source AI from Meta).

• Jevons paradox shows us that efficiency gains tend to lead to greater overall consumption. Good news for hardware owners and data centre operators. Less good news for closed-model providers.

Read the full story on the Register here.

DOJ Moves to Block HPE’s $14B Juniper Acquisition

The Department of Justice is stepping in to block Hewlett Packard Enterprise’s $14B acquisition of Juniper Networks.

The DOJ argues that the acquisition would reduce competition in key networking hardware markets, particularly for enterprise AI and cloud infrastructure. HPE and Juniper have strongly opposed the lawsuit, claiming the deal would strengthen US-based tech leadership against Cisco and Chinese manufacturers. With AI infrastructure scaling at record speeds, the regulatory battle over key networking companies is one to watch.

Why this matters:

• Networking hardware is fast becoming another key battleground for AI infrastructure beyond just compute.

• HPE vs Cisco is already a fight, and adding Juniper would have reshaped the competitive landscape (in favour of US dominance if you buy into the pushback below).

• Given the new US administrations focus on American tech dominance (and deep links to silicon valley), this block may yet be thrown out. However, if the deal collapses, expect both firms to shift strategies fast.

Read the HPE and Juniper Networks response on BusinessWire here.

Lambda Partners with Pegatron for Nvidia GB200 NVL72 Rack Deployments

Lambda is scaling up with Pegatron to deploy Nvidia’s GB200 NVL72 racks.

Lambda has announced a partnership with Pegatron to deploy Nvidia’s GB200 NVL72 racks. These racks integrate 72 Blackwell GPUs and are designed for AI training at hyperscale. With Lambda’s existing focus on providing GPU clusters as a service, this move signals its intent to compete with larger players like CoreWeave and Crusoe, offering an alternative to traditional hyperscalers for enterprises looking to scale AI workloads.

Why this matters:

• Lambda is scaling up with Blackwell GPUs as NeoCloud competition grows.

• Nvidia’s GB200 is already in high demand, and being early to market with a new hardware generation is likely to be highly profitable.

• Coupled with the recent launch of their inference API, it’s clear that Lambda aren’t messing around when it comes to solving customer problems.

Read the full story on Data Center Dynamics here.

Alibaba vs DeepSeek – China’s AI Battle Intensifies

Alibaba has released a new AI model it claims surpasses DeepSeek V3, escalating China’s AI arms race.

Alibaba’s latest AI model reportedly outperforms DeepSeek’s V3 across multiple benchmarks, showcasing the growing competition between China’s biggest tech firms. While DeepSeek’s recent breakthrough in cost-efficient AI development shook global markets, Alibaba’s move suggests it isn’t going to let DeepSeek dominate the Chinese AI scene without a fight. As both companies push heavily into AI model training and infrastructure development, this rivalry is becoming a defining one in China’s push for AI leadership.

Why this matters:

• China’s AI market is splitting into competing ecosystems.

• The move to open source this model as well shows how much dominance

• If Chinese AI companies keep scaling at this pace, they’ll push the rest of the market harder than ever (likely already leading OpenAI to release o3-mini with it’s reasoning capability to free users).

Read the full story on Reuters here.

ASML’s Chip Orders Defy Market Volatility

Despite AI selloffs, ASML’s orders exceeded forecasts proving chip demand isn’t slowing down.

ASML, the Dutch semiconductor giant behind cutting-edge chip manufacturing equipment, reported stronger-than-expected orders, even as AI-related stocks saw turbulence. With demand for high-end AI chips like Nvidia’s H200 and Blackwell still rising, ASML’s ability to sell its extreme ultraviolet (EUV) machines remains a critical indicator of where the semiconductor industry is heading. This comes just as DeepSeek’s efficiency breakthroughs have cast doubt on whether hyperscalers will maintain their breakneck pace of GPU spending.

Why this matters:

• ASML’s EUV machines are essential for next-gen chip manufacturing.

• AI infrastructure investments aren’t slowing, despite stock market volatility.

• Whether this investment trend continues in the wake of DeepSeek remains to be seen.

Read the full story in the WSJ here.

20GW of Natural Gas to Power AI Data Centres

Utilities in the Southeastern US plan over 20GW of new natural gas capacity by 2040 to meet AI data centre demand.

With AI-driven data centres consuming record levels of power, utility providers are shifting their energy strategies. A new report from the IEEFA highlights that over 20GW of natural gas generation will be added across the Southeast, helping to sustain data centre expansion while reducing grid strain.

Why this matters:

• AI data centres are reshaping US energy investment priorities.

• Renewable power can’t scale fast enough, so natural gas is filling the gap.

• Expect more power grid battles as AI’s energy appetite continues growing.

Read the full story on Data Center Dynamics here.

Meta’s AI Infrastructure Spend to Hit ‘Hundreds of Billions’

Zuckerberg says Meta will spend ‘hundreds of billions’ over the long term on AI infrastructure.

Meta’s AI ambitions continue to scale, with CEO Mark Zuckerberg outlining a staggering long-term investment in AI compute. And with a public commitment to deploy 1.3m GPUs before the end of the year, Meta is not playing. While exact spending figures remain unclear, this commitment underscores how hyperscalers are preparing for the next phase of AI growth.

Why this matters:

• Meta is committing at hyperscaler scale AGI-level spending.

• AI workloads demand compute, and Meta’s spending suggests it expects to lead.

• If “hundreds of billions” becomes the new normal, hyperscaler AI budgets are about to hit another level.

Read the full story on Data Center Dynamics here.

The Rundown

If week 4 has taught us anything, it’s that moves to the downside in AI infrastructure can be just as absurd as those to the upside.

DeepSeek’s $6M AI model sent Nvidia into a $589B nosedive, only for it to claw back $260B the next day. Meta pledged ‘hundreds of billions’ to AI infrastructure, while Alibaba hit back at DeepSeek with a model it claims outperforms V3. Lambda locked in Nvidia GB200 racks to stay competitive. ASML smashed chip order forecasts, proving GPU demand isn’t fading (yet), despite fears DeepSeek’s efficiency could change the game. The US is adding 20GW of natural gas to power hyperscale AI, while the DOJ is trying to kill HPE’s $14B Juniper acquisition, setting up a fight over AI networking.

If things carry on like this, I’m going to have to redesign my bingo card.

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.