The New York team is also hiring! We are seeking passionate enthusiasts to join us supporting event planning, sales and sponsorships, and marketing . Apply here!

The AI world seemed to be turned upside down last week when a hedge fund in China called High-Flyer released an open-source model called DeepSeek-R1. Upside down may still be an understatement given a trillion dollars (yes, trillion!) was almost immediately erased from the US technology sector. This model not only showed competitive performance to OpenAI’s most advanced o1 model but was supposedly trained on a couple hundred H100s for under six million dollars! Remember, GPT-4 training costs have been estimated to be $80-100M and GPT-5 is rumored to cost $1B. Even more breathtaking is that this open-source model could be run locally on a high-end computer instead of a $500B network of local GPU-centric datacenters.

If you’re still here, you already know all of this, so I’ll focus on the technology breakthroughs and tangible impact on the AI community. DeepSeek combines novel neural network architectures with innovations such as Group Relative Policy Optimization (GRPO) and Multi-Head Latent Attention (MLA), which has demonstrated the ability to process vast datasets with unprecedented efficiency and cost-effectiveness. These advancements have significant implications for tech giants from Google and Apple to OpenAI and Nvidia reshaping competitive dynamics and AI spending strategies across the AI ecosystem. Let’s dive in!

DeepSeek has achieved results that were once thought unattainable. For example, its R1 model delivers reasoning capabilities comparable to OpenAI’s o1 while reducing training costs by up to 98%, reportedly requiring just $5.6 million for final training​​. Key to this breakthrough are techniques like FP8 Mixed Precision Training and GRPO, which minimize computational overhead while improving performance. This combination has disrupted entrenched paradigms reliant on massive compute resources and energy-intensive infrastructure​​. It’s important to note that there is significant speculation on the authenticity of Deepseek’s claims, but this will surely be validated in the coming weeks.

Nevertheless, by making this model and contributing research open source, DeepSeek has democratized access to the most advanced AI capabilities. Its open-source approach further challenges the proprietary strongholds and compute moats of the most formidable players in the AI ecosystem.

DeepSeek’s compute-efficient architecture that is orders of magnitudes lower than what the industry initially believed was possible offers a pathway for Apple to enhance its privacy-focused AI systems, such as Siri and other applications dependent on on-device inference. Apple’s incredible hardware advantage, including unified memory in Apple Silicon, makes it particularly well-suited to capitalize on DeepSeek’s low-memory inference models. This could bolster Apple’s position in edge AI, where efficient, localized processing is paramount​, and many believe will truly unlock the value of personalized, autonomous systems.

Meta, with its commitment to open-source AI and highest number of consumer touch points, stands to benefit significantly from DeepSeek’s novel approach. They made an early bet and now appear to be in the winning seat. The company could incorporate these techniques into its Llama models, accelerating its dominance in end-user applications and continuing to benefit from the additional developments across their open-source community. That said, the competition DeepSeek introduces into the open-source AI ecosystem will surely force Meta to innovate faster to maintain their developer loyalty​​.

Amazon, and specifically AWS, stands to benefit greatly from DeepSeek’s innovations, as its open-weight models align perfectly with AWS’s strategy of hosting third-party AI solutions. By enabling cost-efficient, scalable AI services, DeepSeek enhances AWS’s value proposition for enterprise customers. With lower costs fueling broader AI adoption, Amazon can expand its dominance in AI infrastructure without the need to invest heavily in proprietary models​​.

As the leader in search and online advertising, Google faces significant risks from DeepSeek. A model that offers near-instant, contextually accurate search capabilities could erode Google’s competitive moat by democratizing access to the most advanced models for use across new search startups like Perplexity. While Google has the resources to adopt DeepSeek’s approach, any disruption to its ad-driven revenue model would necessitate a significant strategic pivot​​. It’s also important to recognize the hundreds of billions of spend into building up their TPU hardware could be seen as less valuable in a world with much more compute efficient models.

DeepSeek directly challenges OpenAI’s business model (relying on highly priced APIs) by offering reasoning models like R1 at a fraction of the training cost and with open weights, undermining OpenAI’s premium pricing strategy. This threatens to commoditize the foundation model layer, reducing the competitive edge of proprietary systems like o1. Their heavy investments in large-scale infrastructure, such as the $500B Stargate initiative, now face questions about sustainability and efficiency in the face of DeepSeek’s leaner approach. While OpenAI may retain an edge in cutting-edge research and consumer products, it must quickly adapt to this new paradigm​​.

Anthropic’s focus on safety-first AI aligns partially with DeepSeek’s transparency in model development, but its proprietary constitutional AI systems face pressure to justify their premium in a landscape where open models can deliver competitive results at a fraction of the cost​​.

Microsoft’s close partnership with OpenAI positions it as both a beneficiary and a potential rival to DeepSeek’s approach. While DeepSeek could complement Microsoft’s enterprise offerings like Copilot by driving inference costs lower while increasing performance (will Copilot finally be helpful! 😛), it also threatens the need for their colossal data center and in-house chip development spend over the last couple of years.

DeepSeek’s efficiency-driven models challenge Nvidia’s dominance by reducing reliance on high-end GPUs like the H100, showcasing that advanced AI can be achieved with less powerful hardware. This threatens Nvidia’s growth narrative, as customers may seek alternative, cost-effective solutions or less compute-intensive architectures. However, increased AI adoption from cheaper models could ultimately drive higher demand for GPUs in inference workloads, keeping Nvidia relevant in a more competitive landscape​​.

DeepSeek’s advancements signal a pivotal moment in AI’s evolution. By challenging the status quo of energy-intensive, high-cost model training, it forces the industry to reconsider its trajectory. Companies best positioned to adapt—like Apple and Meta—may emerge stronger due to their focus on an open-source strategy or on-device capabilities, while those with large foundation models predicated on large compute infrastructure risk losing their competitive moat.

As AI enters this new era of efficiency, the key question remains: Will DeepSeek’s approach become the norm, or will entrenched players adapt quickly enough to neutralize its disruptive potential? The answer could redefine the future of AI development.

🚀 Last night, 300+ AI builders packed the AWS GenAI Loft to witness the next wave of AI innovation with us and Product Hunt! 🚀

Nearly 100 startups applied for a chance to ENTER THE ARENA 🏟️ and the 8 finalists did not disappoint! 🤩

🔥 FEATURED STARTUPS 🔥

📊 Athena Intelligence – AI that thinks like a 24/7 analyst (Brendon)

🔎 Big Smart AI (YC S23) – The best of LLMs + keyword search (Akash)

💡 Gauge (YC S24) – The fastest way to untangle messy codebases (Evan, Caelean)

🤖 Solidroad (YC W25) – Scaling businesses with AI agents (Mark, Patrick)

💻 Codebuff – AI-powered coding, reimagined (James, Brandon)

📊 Cofactory – Next-gen AI for customer acquisition (Miles)

🌍 Nilo Technologies – 3D world creation at record speed (Nuno)

💻 Nuanced (YC W24) - Make AI tools smarter with semantic understanding (Ayman Nadeem)

💥 The energy was electric. The crowd was engaged. And after 30,000+ virtual claps and hundreds of real-time votes, the audience crowned their winners:

🏆 Best Overall: Cofactory

🤖 Best Technology: Big Smart AI (YC S23)

🎨 Most Creative: Nilo Technologies

Huge shoutout to Emergence Capital, Graphite, and Roam for fueling the next wave of AI builders!

🚀 This is only the beginning of what is quickly becoming a global network of events that elevates the entire tech ecosystem worldwide. We intend for this to become the most impactful event series in tech history. THANK YOU ALL for being a part of it!!

We kicked off 2025 with an incredible Research Roundtable featuring thought-provoking discussions on advanced AI agents and algorithmic reasoning. Our distinguished speakers and engaged audience created an atmosphere of deep technical exploration and collaborative learning.

Key Highlights:

The event fostered engaging conversations around:

We're grateful to our speakers and attendees for making this such an enriching experience!

The GenAI Collective is growing rapidly and we’re looking for passionate, visionary community builders to join our team. If you want to join a team of 50+ organizers helping to shape the future of AI, we have tons of exciting ways to get involved! Read more about each opportunity below and learn what you can create with this vibrant community!

About Eric Fett

Eric leads the development of the newsletter and online presence. He is currently an investor at NGP Capital where he focuses on Series A/B investments across enterprise AI, cybersecurity, and industrial technology. He’s passionate about working with early-stage visionaries on their quest to create a better future. When not working, you can find him on a soccer field or at a sushi bar! 🍣