AI Talking Points of Today

March 18, 2025 · Callum Ke

The genuine breakthrough in modern AI—the consumer-facing applications that transform theoretical advances into revolutionary technology. ChatGPT provided this crucial bridge: a simple text interface that democratized access to generative models and reignited widespread interest in artificial general intelligence.

Consumer applications don't merely popularize AI—they rigorously stress-test these systems at unprecedented scale. The demands of millions of diverse, unpredictable use-cases push models to their limits, creates powerful economic incentives for companies to invest in massive infrastructure scaling and continuous model improvements. This virtuous cycle of consumer adoption driving technical advancement has accelerated AI development far beyond what academic research alone could achieve.

This convergence of technical innovation and consumer accessibility has catalyzed transformations across numerous fields: medical diagnostics, accelerating novel research discoveries, creating truly personalized education systems, and revolutionizing compliance, legal, software and financial operations as an intelligent companion for recall and research.

We examine the key trends and players shaping this decade of AI development. A landscape is defined by both technical progress and practical implementation challenges—where theoretical capabilities meet real-world implications that are reshaping the structure of industries, economies, policies and societies worldwide.

Key Players

The global dynamic currently looks like:

• The United States maintains its leadership position in fundamental research, enterprise and consumer AI adoption through its start-up and VC culture.
• China has accelerated its domestic AI ecosystem development through its big corporations and is paving its own way, determined not to be left behind despite export control constraints.
• The European Union has established a distinctive regulatory approach that emphasizes responsible AI development working with the big players around the world.

The US: The Magnificent Seven

Apple, Microsoft, Alphabet, Amazon, Meta, NVIDIA and Tesla continue to dominate the AI landscape in Each has carved out distinctive AI strategies:

• Microsoft leverages its OpenAI partnership to embed AI across enterprise and SaaS offerings while prioritizing sustainable growth trajectories.
• Alphabet dominates benchmarks with Gemini 2.5 Pro while DeepMind and G-Suite integration delivers substantial consumer impact, cementing its position as the R&D powerhouse.
• Meta carves its niche through open-source multimodal foundation models with significant investment, though LLaMA increasingly lags behind competitors.
• Amazon integrates AI across retail, devices, robotics, and developer services. Their competitive advantage lies in proprietary chip development that offers cost-effective training and inference alternatives to NVIDIA.
• NVIDIA maintains unrivaled dominance in high-performance GPUs and data center infrastructure with superior software integration. Their robust three-year pipeline and extensive enterprise partnerships demonstrate overwhelming market demand.
• Tesla repositions as an AI and software company to justify premium valuations despite trailing Waymo and BYD in autonomous driving. Their strategic pivot now focuses on robotics innovation through Optimus.

AI Revenue Projections (2024-2026)

Source: CNBC Tech AI Spending Report, Yahoo Finance AI Market Analysis, Investopedia Magnificent Seven Analysis, Company Earnings Reports

Money Flows

• Combined AI spending for the Magnificent 7 is projected to increase from $290B in 2024 to $488B by 2026.
• Nvidia leads in AI revenue generation, with projected revenue of $130B in 2024, expected to reach $220B by 2026.
• Amazon has the highest AI infrastructure spending, allocating $100B for 2025.
• Microsoft shows the strongest revenue growth trajectory, with AI revenue projected to increase from $40B to $110B between 2024-2026.
• While current spending exceeds revenue for most companies, the gap is expected to narrow by 2026.
• Total AI revenue for the Magnificent 7 is forecasted to reach $615B by 2026, outpacing the $488B in spending.

The Boutique's

The (mostly) privately funded research firms are akin to traditional fashion houses like Gucci, Prada, and Louis Vuitton. The trends set by these luxury brands eventually trickle down and are tweaked to more consumer-accessible retailers, AI research companies follow a similar pattern of innovation and dissemination.

OpenAI, Anthropic, xAI, and now DeepSeek can be viewed as the tone setters of the AI world, each with their own unique flavor. The top three players (OpenAI, xAI, and Anthropic) control over 95% of the total valuation in the general-purpose LLM market.

However, each house cannot rest easily, their cutting-edge techniques are quickly adopted and refined by the research community. As this happens, the demand for what was once considered state-of-the-art (such as Claude-3.5 or DeepSeek 3) falls dramatically. They take significant risks and push the boundaries of state-of-the-art (SOTA) models, continually attracting more capital and resources through private and corporate backing(OpenAI:Microsfot, Anthropic: Google+Amazon) in order to push the AI frontier with the hope to be the first to reach AGI.

General-Purpose LLM Companies by Valuation (2025)

Source: WriterBuddy AI Market Analysis, CNBC Tech Funding Reports, Business Insider, The Information

Chinese Tech Giants

Alibaba, Baidu, Tencent, ByteDance, and SenseTime have developed increasingly sophisticated AI capabilities despite navigating a complex regulatory environment both domestically and internationally. China has already caught up on foundational model development.

• Alibaba: QWEN
• Baidu: Ernie
• ByteDance: DuoBao
• Tencent: Hunyuan

Chinese LLMs vs ChatGPT 4o

Data centers and GPU's, please.

The global distribution of AI computational infrastructure investment and research continues to reflect market demand globally. Data center capacity has expanded significantly in key regions including North America, Europe, and Asia-Pacific, with particular growth in infrastructure optimized for AI (and non-AI) workloads. Yes, data-centers are not just for AI (looking at you Cloud).

Global AI Data Center Distribution (2025-2030)

Year: 2025

Source: JLL's 2025 Global Data Center Outlook, ABI Research Data Center Projections, Goldman Sachs Research, Deloitte Data Center Electricity Consumption Predictions

Manufacturing Concentration: Chip production remains largely controlled by a small set of companies:

• Taiwan's TSMC (54% of global foundry revenue) produces chips for NVIDIA, AMD, and Intel
• The Netherlands' ASML holds a monopoly on advanced lithography machines essential for cutting-edge semiconductor manufacturing
• Samsung (South Korea) is emerging as a secondary production hub for GPUs

Export Controls and the NVIDIA Problem

Since 2022, the United States has implemented progressively stricter controls on advanced semiconductor exports to China, particularly targeting high-performance GPUs:

• October 2022: Initial controls focusing on advanced computing chips and semiconductor manufacturing equipment
• October 2023: Expanded controls lowering performance thresholds and closing loopholes, affecting Nvidia H800 and A800 chips previously designed for the Chinese market
• 2024: Further refinements targeting emerging technologies and expanded end-use restrictions

NVIDIA GPU Comparison: Export-Controlled vs. Chinese Market Models

Feature	H100 (Controlled)	H800 (Initially Allowed, Now Restricted)	H20 (China Market)	L20 PCIe (China Market)
FP8 Compute (Training)	3,958 TFLOPS	1,979 TFLOPS	296 TFLOPS	191 TFLOPS
HBM Memory	80GB HBM3	80GB HBM3	96GB HBM3e	48GB GDDR6
Memory Bandwidth	3.35 TB/s	3.35 TB/s	2.86 TB/s	768 GB/s
Chip-to-Chip Interconnect	900 GB/s NVLink	400 GB/s (Restricted)	250 GB/s	Limited NVLink
Training Efficiency	100%	~50%	~20%	~15%
Export Status	Fully Restricted	Initially Allowed, Restricted since Oct 2023	Currently Allowed	Currently Allowed

Source: Nvidia specifications, SemiAnalysis reports, U.S. Department of Commerce export regulations

As Dario Amodei argues(CEO Anthropic): "Making AI that is smarter than almost all humans at almost all things will require millions of chips, tens of billions of dollars (at least), and is most likely to happen in 2026-2027." This timeline creates urgency for both sides of the export control debate.

• Huawei is scaling production of its Ascend AI processors (planning 100,000 Ascend 910C and 300,000 910B chips in 2025)
• Chinese companies like Biren Tech(壁仞科技) and Moore Threads(摩尔线程) have started created their own GPU architecture (BR100 and MTTS4000)
• Shenzhen SiCarrier (新凯来 Huawei-backed) has unveiled an extensive range of chipmaking tools at SEMICON 2025, potentially challenging ASML's dominance
• SMIC (中芯国际 - Semiconductor Manufacturing International Corporation, Huawei backed) is targeting 5nm process technology by 2027, though access to EUV lithography remains a significant hurdle
• Despite restrictions, China accounted for 36% of ASML's total revenue in 2024, though this is projected to drop as tighter controls take effect. Chinese companies are accelerating purchases of older-generation DUV (Deep Ultraviolet) lithography tools in response.

The Rise of Reasoning

AI models have evolved rapidly from pattern recognition, to generation to sophisticated reasoning capabilities. OpenAI opened the doors with GPT-o1 while DeepSeek raised the sealing with DeepSeek-R1.

Reasoning enables the ability to process information through multiple logical steps(sequential), evaluate evidence, and generate novel insights.

• Multi-step logical reasoning through complex problems
• Generating and evaluating multiple solution paths
• Incorporating external tools and knowledge sources
• Explaining decision processes in human-understandable terms

The Cost of Reasoning

Despite remarkable progress, several critical bottlenecks constrain both the development and deployment of reasoning-capable AI systems:

• Computational Efficiency: Reasoning operations require significantly more compute resources than traditional inference, with corresponding increases in latency and cost
• Memory Bandwidth: Advanced reasoning requires maintaining and manipulating complex state information, creating memory bottlenecks in current hardware architectures
• Energy Consumption: The energy requirements for reasoning operations grow non-linearly with model complexity, creating sustainability challenges
• Specialized Hardware: Current accelerators optimized for matrix multiplication operations are suboptimal for the sparse, conditional computation patterns of reasoning workflows. Examples include: dataflow architectures IPU's, IBM TrueNorth, dynamic-flow ASIC/FPGA's and hybrid CPU-GPU models.

The Critical Role of HBM Memory

High Bandwidth Memory (HBM) has emerged as a critical component for advanced AI systems, particularly those focused on reasoning capabilities. This specialized memory architecture addresses the significant bandwidth requirements of complex AI workloads by stacking memory dies vertically and connecting them with through-silicon vias (TSVs).

Key advantages of HBM over traditional GDDR memory include:

• Bandwidth Improvements: HBM offers 3-7x greater memory bandwidth than GDDR6X, enabling more efficient data access during complex reasoning operations
• Energy Efficiency: HBM consumes approximately 50% less power per bit transferred compared to GDDR alternatives
• Spatial Efficiency: The stacked die architecture allows for greater memory capacity in a smaller footprint, critical for data center density
• Context Window Support: HBM's bandwidth and capacity are essential for supporting the extended context windows that enable sophisticated reasoning across long documents or multiple sources

Leading HBM Manufacturers

HBM production capacity has become a determinant of AI accelerator availability, with manufacturers operating at full capacity and commanding premium pricing

• SK Hynix: The current market leader, supplying HBM3E memory for NVIDIA's H200 and upcoming Blackwell GPUs
• Samsung Electronics: A major player focused on scaling HBM4 production with innovations in TSV density
• Micron Technology: Rapidly gaining market share through aggressive performance improvements and manufacturing expansion
• CXMT (ChangXin Memory Technologies): The leading Chinese memory manufacturer developing domestic HBM alternatives amid export restrictions

Application Case Study: Revolution in Software Development

The software development landscape is undergoing a profound transformation driven by AI, creating both significant opportunities for productivity gains and challenging established workflows as knowledge is democratised. This shift represents one of the most immediate and tangible applications of AI capabilities and often used as a benchmark for performance.

AI-powered coding assistants have rapidly evolved from simple autocomplete tools to sophisticated pair programmers capable of understanding complex codebases and generating production-ready code:

Developer to Designer

Beyond individual productivity tools, AI is fundamentally changing how software is conceptualized, designed, and maintained. Using Amazon as an example, top-down efforts are being sued to encourange and educate developers to use AI tools across all parts of the development life-cycle.

The most transformative impact may be in how AI tools handle increasing system complexity, allowing developers to reason at higher levels of abstraction where humans specialise in while AI manages implementation details. Contrary to early concerns about job displacement, evidence suggests AI tools are primarily augmenting rather than replacing developers, with demand for software development talent continuing to outpace supply in most markets.

• Skill Evolution: Premium skills are shifting from syntax mastery to effective AI collaboration, architecture design, and prompt engineering
• Productivity Amplification: Organizations report 20-40% productivity gains for experienced developers working with AI assistants on appropriate tasks
• Entry Barrier Reduction: AI tools are making software development more accessible to those with limited formal programming education
• Specialization Changes: Previously distinct specializations are blurring as AI tools enable developers to work effectively across multiple domains

In Closing

It's 2025, and the AI industry stands at a pivotal moment. Foundation model development has accelerated at a breathtaking pace, with each new iteration surpassing its predecessor in rapid succession. This acceleration has fundamentally reshaped the global landscape to accommodate the current and future demands of artificial intelligence.

Yet the industry treads a precarious path. Questions arise about whether we're building supply infrastructure for demand that may never materialize. For example, Microsoft CEO Satya Nadella recently expressed skepticism regarding massive investments without clear end goals, signaling a shift toward prioritizing sustainable growth trajectories rather than unchecked expansion.

Undeniably, a growing undercurrent of fear and resentment is emerging—both in public sentiment and tangible reality—that AI is diminishing human agency and creativity, potentially fostering laziness, and lack of checks towards outcomes that yield utility.

We must take a step back and ask ourselves: Is AI's metaphorical flower still in the process of blooming, yet to reach its full splendor? Or has it fully blossomed, ready to be harvested so consumers can reap its rewards? Alternatively, has its growth been stunted by harsh environmental conditions, requiring us to await the next favorable season? The gardener—representing collective human agency and decision-making—has yet to determine the answer.