In partnership with

President Donald Trump, flanked by tech industry leaders, announced a significant investment in artificial intelligence on January 21, 2025: Project Stargate.

While the announcement came during the Trump administration, the project itself predates his presidency, with origins in the Biden administration.

This ambitious project, a joint venture between OpenAI, Softbank, MGX, and Oracle, aims to invest $500 billion over the next four years in constructing data centers and the necessary energy infrastructure to power the next generation of AI systems in the United States.

This initiative signals a significant step towards securing American leadership in AI, potentially creating hundreds of thousands of jobs and generating substantial economic benefits.

Then DeepSeek came on the scene and the U.S. is doubling down, as I covered last week, the fear that China was beating us has many folks not in the know running scared.

I’m not. Even though DeepSeek was very innovative these models are all going to benefit from their work. I think the takeaway from this is that the race for capable models that achieve AGI(artificial general intelligence) will finish roughly at the same time. This is an example of multiple discovery (also known as simultaneous invention) a well-known scientific hypothesis in new technology breakthroughs.

However, the reality is that we’ll have an insatiable hunger for AI and the infrastructure to run it for the foreseeable future — that’s called Javon’s Paradox for those geeks among us.

Read on for this primer on the infrastructure we’ll need to keep things running.

OpenAI's $500 Billion Project Stargate

How can a 100 Billion dollar company make a $500 Billion Investment in AI Infrastructure?

Technological revolutions often push existing infrastructure to its breaking point, forcing rapid innovation. The early internet exemplifies this. In the mid-1990s, most users connected through dial-up modems operating over standard phone lines. At 56 kilobits per second (kbit/s), downloading a single 5MB file could take nearly 15 minutes. For context, modern broadband speeds in the U.S. average 200 megabits per second (Mbit/s)—over 3,500 times faster. What took minutes back then is now nearly instantaneous.

Behind the scenes, the National Science Foundation Network (NSFNET), which operated at just 56 kbit/s in 1986, quickly became overwhelmed as the Internet’s popularity exploded. Upgrades to T-3 lines (45 Mbit/s) provided temporary relief, but demand consistently outpaced supply. Bottlenecks emerged at peering points like the Commercial Internet Exchange (CIX), where data from various networks converged, creating significant delays. Early Internet users had to wait, unable to stream video or even load basic images without frustration.

Today, artificial intelligence (AI) faces a similar infrastructure challenge. Modern AI workloads, particularly the training and inference of large models, are testing the limits of current hardware and data center designs. These workloads require exponentially more computational power, memory bandwidth, and energy.

Artificial intelligence (AI) applications require GPUs and AI hardware accelerators that can efficiently process data-intensive and computation-intensive workloads. We already are seeing a shortage of GPUS. AI accelerators require two types of memory: weight memory, which stores the parameters of the AI models, and buffer memory, which stores the intermediate input or output data when computing a portion of the AI models.

This means that memory could be the next bottleneck in the AI supply chain. Memory needs during training can reach three to four times the final model size because intermediate activations must remain in memory. This strain has led engineers to explore new chip designs that place memory closer to processing units or use optical-based methods.

The urgency is evident in growth rates. Researchers at the University of California, Berkeley report that while transformer models have grown 240 times every two years, AI hardware memory has only doubled over the same period. Over the past two decades, peak server FLOPS have risen by a factor of three every two years, but DRAM and interconnect bandwidths have scaled by only 1.6 and 1.4, respectively. Bridging this gap requires hardware that can handle the rapidly rising computational load without stalling on data transfer.

Without intervention, these bottlenecks could stall AI progress, just as inadequate bandwidth once constrained the early Internet.

Enter Project Stargate

Project Stargate represents a transformative shift in how AI infrastructure is designed and implemented. Unlike traditional cloud data centers, which are optimized for generalized workloads, Stargate focuses on facilities tailored specifically for AI. These centers aim to resolve two key issues: computational power and energy constraints.

AI models require massive numbers of interconnected processors capable of handling parallel computations. However, existing facilities often face bandwidth limitations and power capacity shortfalls. Stargate’s solution includes building advanced data centers featuring cutting-edge chips and localized energy generation, such as solar and wind farms, to ensure a stable power supply.

AI infrastructure is expensive, and adoption has been incentivized through aggressive pricing strategies. Last year, at this time, OpenAI Altman was in the news for his assertion that we needed to invest $9 trillion in AI infrastructure to meet the demand. This is nothing new, we had the same issues when we built out the Internet. However, the AI growth curve as hard as it may be to believe is going to be faster.

Stargate, supported by OpenAI, SoftBank, and the Middle East AI fund MGX, plans to invest about $100 billion, potentially exceeding $500 billion, into data centers for OpenAI’s AI workloads. According to the report, SoftBank and OpenAI will initially contribute around $15 billion each. The aim is to gather equity from current investors and secure debt to achieve this ambitious project.

This week, Google/Alphabet announced a substantial increase in its capital expenditures for 2025, planning to invest $75 billion—a 43% rise from the previous year. The investment will primarily target AI infrastructure, including servers, data centers, and networking equipment, to support the growing demands of artificial intelligence applications. This move aligns with similar commitments from other tech giants, such as Meta and Microsoft, who have also pledged significant investments in AI infrastructure.

Not to be left out. Meta Platforms announced plans to invest between $60 billion and $65 billion in 2025 to enhance its artificial intelligence infrastructure. This investment will focus on expanding data centers and increasing GPU capacity to support AI initiatives. CEO Mark Zuckerberg emphasized the company's commitment to AI, stating that he expects Meta AI to become a leading assistant used by over a billion people.

In addition to infrastructure expansion, Meta is constructing a $10 billion AI data center in northeast Louisiana, marking its largest facility to date. The company also plans to deploy over 1.3 million GPUs by the end of the year to support its AI operations.

These numbers are unprecedented but probably even underfunded should AI reach these new heights of infrastructure utilization.

The Subsidy Factor

OpenAI’s ChatGPT Plus, which costs $20 per month, is likely being subsidized to attract more users. The price is low when compared to the operational costs of running large-scale inference models. Meanwhile, OpenAI’s ChatGPT Pro, priced at nearly $200 per month, offers significantly higher value for power users but even then it’s not clear if that’s profitable or a loss-leader. This disparity underscores a broader strategy: subsidize entry-level services to encourage widespread adoption, while premium tiers cover the true costs.

This approach mirrors early internet service providers, which offered low-cost plans to drive adoption while investing heavily in infrastructure. However, AI companies face unique challenges due to the sheer cost of training and deploying large models. Training a single model like GPT-4 can cost tens of millions of dollars, with inference costs compounding as user demand scales. Without subsidies, widespread adoption of AI tools would likely slow considerably.

Beyond the way frontier model providers like Meta, OpenAI, Google, and Anthropic are spending ahead of revenue, the U.S. government is also doing the same. The CHIPS and Science Act, signed into law in August 2022, aims to bolster semiconductor manufacturing and research in the United States. It allocates $52.7 billion in funding for domestic chip facilities, including grants, loans, and loan guarantees for companies that establish or modernize manufacturing plants. The legislation also provides a 25% investment tax credit, incentivizing greater private-sector involvement and addressing a critical need for secure, resilient supply chains.

Beyond direct funding, the law prioritizes research and workforce development to meet advanced technology demands. Agencies such as the Department of Commerce, the National Science Foundation, and the Department of Energy receive additional resources to support microelectronics breakthroughs, next-generation AI hardware, and specialized training initiatives. The intent is to reduce reliance on overseas suppliers, foster innovation, and ensure the domestic workforce can support and sustain emerging chip technologies.

Memory Bandwidth and Alternative Hardware

Even with initiatives like Stargate, there are other bottlenecks. For example, memory bandwidth remains a critical bottleneck. AI workloads demand processors that can rapidly exchange vast amounts of data, but existing memory architectures haven’t kept pace. High-bandwidth memory (HBM) systems, such as HBM3, are helping address this gap but come at a significant cost.

To tackle these challenges with infrastructure not just memory but the need for GPUs is unheralded. With NVIDIA GPUs in high demand, companies are exploring alternative processors for AI training and inference:

• Graphcore IPUs: Designed for parallelism, IPUs excel at tasks requiring high computational density, such as training large language models.

• Cerebras CS systems: Built around wafer-scale engines, Cerebras processors offer unparalleled memory bandwidth for training massive AI models.

• Tenstorrent: A rising player in AI hardware, Tenstorrent focuses on flexible chip designs that optimize both training and inference workloads.

• AWS Inferentia: Amazon’s custom inference chips are tailored to reduce costs and energy consumption for deploying AI models at scale.

• Google TPUs: Tensor Processing Units, now in their fifth generation, offer specialized hardware optimized for both training and inference in Google’s ecosystem.

Each of these alternatives offers unique advantages, providing companies with options to tailor their AI infrastructure based on workload requirements and cost considerations.

What Companies Should Consider

For businesses integrating AI, the current infrastructure challenges carry significant implications. Here are key considerations:

1. Evaluate Long-Term Costs: Subsidized tools like ChatGPT Plus may offer short-term affordability, but scaling AI within an enterprise requires planning for the full cost of training, inference, and infrastructure.

   Example: Organizations heavily reliant on large language models (LLMs) must analyze the total cost of ownership (TCO), factoring in energy and hardware investments.

2. Invest in Specialized Hardware: Traditional CPUs and GPUs may not suffice for cutting-edge workloads. Explore alternative processors that offer better efficiency for AI tasks.

   Resource: Graphcore’s IPU offers high-performance AI processing tailored for NLP and computer vision tasks.

3. Adopt Sustainable Practices: AI is energy-intensive. Localized renewable energy solutions, such as those planned for Stargate, can reduce both costs and carbon footprints.

   Case Study: Google’s AI operations aims to use carbon-neutral data centers.

4. Diversify Partnerships: As AI infrastructure evolves, reliance on a single provider can limit flexibility and innovation. Collaborate with multiple vendors to ensure scalability and resilience.

   Example: OpenAI’s move away from exclusive reliance on Microsoft’s Azure to partnerships with Stargate demonstrates the benefits of diversification.

5. Future-Proof AI Investments: AI workloads are scaling rapidly. Companies must prepare for future demands by investing in infrastructure capable of handling next-generation models especially ones that have significant memory, power, and networking needs.

   Forecast: AI models are doubling in size every 6-12 months, necessitating significant hardware upgrades to remain competitive.

A Critical Crossroads

The parallels between the early internet and AI are unmistakable. Both revolutions introduced unprecedented demands on infrastructure, forcing innovation to keep pace. Projects like Stargate represent a new era of AI infrastructure, designed to overcome current bottlenecks in computation, memory, and energy.

For companies, these developments present both challenges and opportunities. The subsidized adoption of AI tools, while helpful in the short term, underscores the need for careful planning around long-term infrastructure investments. The decisions businesses make today—about hardware, energy use, and vendor partnerships—will define their ability to compete in the AI-driven economy of tomorrow.

As history has shown, technological revolutions are ultimately won by those who invest in the right foundations. Whether Stargate opens the portal to AI’s future—or leaves us grappling with bottlenecks—remains to be seen.

Seeking impartial news? Meet 1440.

Every day, 3.5 million readers turn to 1440 for their factual news. We sift through 100+ sources to bring you a complete summary of politics, global events, business, and culture, all in a brief 5-minute email. Enjoy an impartial news experience.

Join for free today!

OpenAI Deep Research - Deep research is OpenAI's next agent that can do work for you independently—you give it a prompt, and ChatGPT will find, analyze, and synthesize hundreds of online sources to create a comprehensive report at the level of a research analyst. Powered by a version of the upcoming OpenAI o3 model that’s optimized for web browsing and data analysis, it leverages reasoning to search, interpret, and analyze massive amounts of text, images, and PDFs on the internet, pivoting as needed in reaction to information it encounters.

Inbox Pilot - Connect your email, train AI with your company data, and let it manage all incoming emails around the clock—delivering responses in your brand's voice and with your expertise.

Lecca.io - Build and customize AI agents with your preferred LLM models. Give your agents superpowers with an expanding toolkit - from web search and phone calling to Gmail and calendar management. Bring your own API keys or use platform credits, it's your choice.

Grok Browser Extension - Grok Button brings the power of Grok to your browser via an extension. Ask Grok a question by typing "grok {your question}" directly in the browser bar.

Breaking (and Correcting ) Bad Writing Habits

We all develop habits—some help us, others hold us back. AI, on the other hand, operates differently. Unlike us, it doesn’t form habits; it generates responses probabilistically, meaning even the same input can yield different results. This is both a blessing and a curse.

When reviewing your writing, you may unknowingly fall into patterns that weaken clarity, structure, or persuasiveness. Instead of relying on the same approach every time, use AI to challenge assumptions, spot redundancy, and refine your message.

Use the prompt below to analyze any piece of writing for clarity, coherence, and impact. I use it to iterate on something I already generated in ChatGPT to make it better (usually).

How to Use This Prompt

You can upload or cut and paste your writing into the prompt at the end. Or you can execute the prompt them cut and paste or upload a document.

# Writing Evaluation & Refinement Prompt  

## Objective  
To analyze a piece of writing for clarity, structure, and effectiveness. The goal is to identify areas where the text could be improved by reducing redundancy, enhancing readability, and strengthening its impact.  

## Role  
You are an expert writing assistant with a deep understanding of effective communication. Your task is to critically evaluate the provided text, offering constructive feedback and specific recommendations for improvement.  

## Instructions  
- **Clarity & Coherence:** Identify sentences or sections that may be unclear, vague, or overly complex. Suggest revisions that improve readability.  
- **Conciseness:** Highlight redundant or wordy phrases and propose more direct alternatives.  
- **Structure & Flow:** Assess the logical flow of ideas. Suggest reorganizations if necessary to improve readability and engagement.  
- **Persuasiveness & Impact:** Evaluate the strength of arguments or messaging. Offer ways to make the writing more compelling.  
- **Grammar & Style:** Point out any grammatical inconsistencies or stylistic weaknesses and provide corrections.  

## Expected Output  
- A **summary of key issues** found in the text.  
- **Specific suggestions** for rewording, restructuring, or improving sections.  
- A **readability and persuasiveness score (1-10)** with a short explanation of the rating.  

I appreciate your support.