Welcome to the Halftime Show.

For our 15th edition, we’re doing something special: a State of the Union on AI as of April 2026.

We’ll skip the technical benchmark deep dives that you can find elsewhere. Instead, we’re going where most don’t: Macro, Infrastructural, and Commercial.

This is a long one and deviates from our regular Thursday format.

Grab a drink, sit back, and enjoy- it’s time to lock the fk in.

On June 13, 2019, the Toronto Raptors beat Golden State in Game 6 to win their first NBA championship. The greatest upset in NBA history (according to every Canadian). I was there. I'll remember that night for the rest of my life.

How did Raptors, the underdog from the North of the border, made up of Lowry, undervalued his entire career; Kawhi, a one year import from South Cal; Pascal, a kid from Cameroon who didn't start playing ball 'til 16; Ibaka, traded twice and written off; Van Leet, undrafted and unsigned out of college; and Gasol, dubbed the 34 yr old "basketball dinosaur" END the GOLDEN STATE WARRIOR dynasty?

Surely not with the most talent or the biggest budget.

Masai Ujiri, the president, understood the dogma. You don't need five generational talents. You need the right pieces- trades, player development, coaching hires, switchable defensive -  assembled into a system where every part makes the others better.

You wouldn't answer "which is the best team in the NBA" with "Luka Dončić." Yet that's exactly how we evaluate AI- by asking who has the best model… a profound category error.

On December 3rd in Washington DC, Jensen Huang shared his "five layer cake" analogy for AI: Energy, Chips, Infrastructure, Models, and Applications.

We'll use that framework to evaluate where the most significant technological shift in human history is heading - in true Delta style, as if you're the GM scouting a championship franchise.

What we’re gaming out:

By the end, you'll understand "AI" better than most in the room.

In any elite sports program, nutrition and conditioning is the unsexy, invisible work that no one watches- but it's what allows you to go harder, faster, longer.

Energy is the conditioning program of the AI economy. Every calculation performed by a model, every chip, every data demands electricity. Without it in abundance and at economic cost, the rest of the stack can't scale.

It is also the primary bottleneck. Energy is a finite organic resources that takes time to harvest, generate, and scale. In systems terms, it's the "limiting factor" - the slowest component to scale determines the ceiling for everything above it.

Currently, the global energy landscape reveals a structural imbalance in capacity and cost, driven by aggressive government subsidies:

Even more granular:

At the Nvidia GTC last month, Jensen Huang rebranded data centers as "AI factories" (more on this later). The product is tokens, and the primary input cost is energy.

Energy cost differential creates a strategic tradeoff: inferior chips become viable through volume deployment, changing the game from 'best chip' to 'most system throughput per dollar.

A chip that draws more power than Nvidia's is a liability in high cost energy markets-but becomes viable at volume when energy is subsidized.

Deploying 4x more chips compensates for per unit efficiency gaps and enables faster infrastructure buildout: more fabs, more data centers, more models, more applications.

Real example: Huawei's Ascend 910C delivers 80% of H100 computing power, consumes significantly more energy per operation, remains economically viable due to subsidized power.

In 2025, China invested $500B in energy projects, an 11% increase YoY.

This is industrial capacity expansion designed to remove energy as a constraint on AI.

For context: China added more renewable energy capacity in 2025 than the entire U.S. grid added from all sources. The 15th Five Year Plan (2026–2030) projects 200–300 GW of new renewable capacity annually-sustaining this pace for the next FIVE years.

Energy keeps the lights on. Chips are the training ground, the weight rack, the tools that turn raw capacity into performance. The game is system level throughput: how much compute you can deploy at what total cost.

But if the best training facility in the world bans you from entry, what do you do?

You build your own gym. Not an Equinox or Gold's at first, but you run it 24/7, and figure out how to make ten athletes training together perform like one elite unit.

Nvidia, as we all know, obliterates the semiconductor layer in per chip efficiency. Nvidia also owns the software ecosystem built around them. CUDA has twenty years and hundreds of millions of installations. Switching costs compound with every model trained.

The H100 and Blackwell architectures set the performance standard globally, and export controls were designed to keep it that way.

The assumption was largely wrong.

Export controls didn't halt development- they incubated Nvidia's challenger through vertical integration at speed (we wrote about how).

When you can't access EUV lithography for 3nm chips, you architect around the constraint. And the current answer?

The SuperPod: thousands of lower performance chips interconnected to function as a single logical machine, making up for what each other lacks.

In technical terms, Huawei's Atlas 950 binds 8,192 Ascend chips with 16 PB/s interconnect bandwidth (10x global internet peak), creating a unified system that learns, thinks, and reasons as one. Like a mini data centre.

A type of architectural innovation that turns volume into a moat.

Production trajectories demonstrate a rapid closing of the gap:

The capital is flowing to volume production, not frontier research. 19 semiconductor firms initiated Hong Kong IPOs in 2025, with 30 more in A share pipeline targeting $13.6B combined fundraising.

We watch both with equal excitement.

All said and done, silicon capability remains theoretical without the physical velocity to deploy it at scale. That's where infrastructure enters.

You can have the best QB and the perfect playbook. But if your coaching staff can't install the system faster than the opponent adjusts their defense, you lose.

Infrastructure is how fast you go from blueprint to deployment. Data centers. Power grids. Manufacturing lines. Supply chains. The physical foundation that determines whether your AI runs in a lab or at scale across millions of users.

Speed of installation beats brilliance of design, especially when AI democratizes output quality.

China's deployment velocity is the no huddle offense of global tech. This phenomenon is what we're most interested in at LOD.

Without going too deep, three decades of supply chain concentration built the Pearl River Delta into the densest hardware ecosystem on earth. Every component supplier, assembler, and testing facility operates within a 100 kilometer radius. Vertically integrated campuses and streamlined permitting accelerate execution further.

The result is the infamous "72 hr prototype to production" cycle.

Non exhaustive, but indicative illustration below-

The humanoid robotics sector illustrates this tempo at its best:

(Watch this)

The infrastructure layer determines whether silicon capability remains theoretical or becomes deployed at billion user scale.

This is something you truly have to see with your own eyes to believe. It's what we witness everyday, here in Shenzhen.

Now that you have the conditioning, equipment, and coaching dialled in, the talent is ready. AI models are the players. The talent you develop ontop.

They get trained, refined, benchmarked, scouted by enterprises, deployed into production.

Different strengths, different specializations, different cost profiles. Some are franchise stars with massive contracts. Others are high value free agents available to anyone willing to sign them. And some, are unassuming rookies who will become generational talent.

All layers matter, but models are where most leverage is built.

The US leads in frontier model development, bar none. OpenAI's GPT-5.2, Anthropic's Claude Opus 4.6, Google's Gemini 3- these set the performance ceiling. They're also closed and controlled.

When U.S. export controls landed in October 2022, cutting off access to Nvidia's H100 and A100 chips, China faced a choice: wait for domestic chip alternatives to catch up, or architect around the constraint entirely. They chose the latter.

Open source became a national ideology- not out of charity, but necessity.

The AI+ Initiative (August 2025) made it official: government subsidies for open source development, explicit directive to "promote open source AI ecosystems." Today, more than 90% of Chinese enterprises use open source technologies.

Closed source approach:

Open weight approach:

What happened next…

Side note, this isn't just enterprise. Take a look at this video created by ByteDance's Seedance 2.0 video gen tool, that caused cease & desist letters from Disney, Paramount, Sony, Warner Bros., Netflix, and MPA within 24 hours of release.

(Watch this)

Incredible.

When DeepSeek R1 was trained for ~$6 million, matching GPT-4-class performance at 1/20th the cost, it proved frontier level intelligence could be commoditized.

Within six months, a wave of models non-Chinese developers had never heard of debuted to market. Moonshot's Kimi K2.5. Alibaba's Qwen. MiniMax. Zhipu's GLM. All open weight & priced near cost. By November 2025, Chinese models held 7 of the top 10 download spots on Hugging Face.

Across the broader landscape:

The adoption numbers tells us whether open source as a distribution tactic worked:

Each derivative represents engineering hours, training data, and product architecture invested into the ecosystem. Switching costs accumulate with every deployment.

And finally, the validation from the big boys:

The cost advantage unlocked application depth that wasn't economically viable before.

When you can run the same intelligence for 3-5% of the price, integration becomes pervasive rather than selective.

By late 2025, the model layer is starting to flatten. Competitive advantage migrated to deployment and execution.

Back to the Raptors. Championships are what you play for. Everything else- the training, the trades, the system- exists to win rings.

That's the Applications layer. Everything built in the layers below…energy, chips, infrastructure, models….only matters if it translates into applications people actually use. Applications are the only layer that generates revenue.

It’s also the only layer that creates the data flywheel feeding everything below it:

In the rest of the world, AI is largely a destination -you go to ChatGPT, you open a separate tool.

In China, AI is plumbing. It runs inside ecosystems that already command hundreds of millions of users, in what's known as "super apps".

The "Super App" Phenomenon

Super apps exist in China due to a combination of technological, economic, social, and regulatory vectors… but most notably, China leapfrogged the PC era completely, and came online via smartphones, creating demand for a single platform that handled everything rather than juggling multiple specialized apps.

Vs the rest of the world who were already locked into specialized, OS level, fragmented apps.

For example:

The Aggregate: 2.2 billion monthly active users across the top 100 AI companies in China. Six of the ten largest AI companies by user scale are Chinese:

Simply put, super apps capture the entire lifecycle of human intent - generating training data across the full user journey that silo'd apps can't replicate.

Going Global

Of the top 23 Chinese AI companies by revenue, 19 generate the majority of their income overseas.

The playbook = "chuhai" (going to sea) = dominate domestically for scale and iteration speed, then export to dollar denominated markets for margin.

Winning a championship is hard. Building a dynasty is harder. The difference is whether you've built a system that compounds.

While the first half of the presentation solidified NVIDIA’s dominance in data centers and agentic software, a massive 45 minute segment was dedicated to what Huang calls the "next trillion-dollar wave": Physical AI.

Actually, he put Olaf to work as a demo.

(Watch this)

The era of the chatbot is maturing; the era of Embodied Intelligence is just beginning.

Jensen is betting the roadmap on AI that moves- robotics, autonomous vehicles, and industrial scale deployment:

Deploying intelligence in the real world requires a different toolkit than building a LLM:

This is where the two loops described earlier converge with accelerating force:

*Restricting access to cutting edge training chips constrains Loop One (frontier model development). It does not constrain Loop Two (physical data accumulation)

A factory inspection model doesn't need a cluster of H100s; it needs millions of labeled production images. The entity that deploys 10,000 robots owns that dataset.

The Bottom Line: Physical AI is the convergence of digital intelligence and industrial muscle. Deployment is the new Moat.