Top Tech AI Stocks 🤖 🔥 📈

For anyone doubting it, here's the three-page Goldman Sachs report on $NBIS. The upgrade was driven by higher sales estimates and an unchanged 7x CY27E EV/Sales multiple. Lowered 2026 Revenue estimates from $6.6B to $3.8B. Raised 2027 Revenue estimates from $5.8B to $7.7B.

$NBIS just released a whitepaper on "The Economics of AI Clusters." 🧐 This paper is a clear breakdown of what actually drives the cost structure of foundation model training, and how $NBIS' vertically integrated AI cloud architecture is designed to make that process dramatically more efficient. Here are the key takeaways: 1) AI workloads are massively expensive, but most inefficiency is hidden. AI infrastructure takes up the largest share of total project cost. GPUs and accelerators deliver huge performance gains over CPUs but come with steep energy, networking, and cooling requirements. In $NBIS' example, training a large language model on a 3,000-GPU cluster costs about $144k per day, and can easily reach several million dollars over a multi-week training cycle. But the real problem isn't just price, it's utilization. Average GPU utilization (MFU) for large-scale training hovers around 45-55%, meaning nearly half of GPU capacity sits idle due to code inefficiencies, interruptions, or system limits. 2) $NBIS quantifies the problem, and how it fixes it. Using a 3,000-GPU training example, the whitepaper compares Nebius to Cloud X, a “fictional baseline GPU provider operating at industry-standard reliability.” On Cloud X, the 3-day theoretical training job stretches to ~190 hours (8 days) once failures, checkpointing, and setup time are factored in. Nebius’ infrastructure, built for AI from the ground up, reduces that to 152 hours, saving roughly 40 hours per job (a 21% improvement). How? Because its clusters are more reliable and recover far faster when failures occur. $NBIS' systems show an average 33-hour mean time between failures (MTBF) and a 12-minute mean time to recovery (MTTR), versus roughly 9.8 hours and 1 hour, respectively, for standard setups. These improvements cut downtime and wasted computation, leading to faster completion and higher GPU utilization. 3) Higher hourly prices, but lower total costs. Even with a higher per-GPU-hour price ($3.50 vs. $2.80 for Cloud X), $NBIS achieves nearly identical total job costs, and its efficiency translates into roughly 25% higher GPU cost efficiency, freeing up about $400k in GPU value on a single 3,000-GPU training run. Example (NVIDIA HGX H200): • Total training time: 189.8h vs 152.1h • Training job cost: both ~$1.59 M • Saved GPU time: 2.2h vs 39.9h • Saved GPU cost: $18k vs $419k This assumes a company reserved an 8-day (192-hour) 3,000-GPU cluster for the job. That “saved GPU cost” represents the value of idle compute avoided, efficiency that translates into roughly 25% higher GPU cost efficiency for $NBIS. For large AI labs running multiple foundation-model training cycles per year, that advantage compounds into millions of dollars in annual savings, even when total job costs look nearly identical. 4) Beyond hardware: operational savings compound the advantage. $NBIS layers multiple cost-saving features on top of its core infrastructure: • Free expert support from in-house AI engineers (no consulting fees) • Managed orchestration tool Soperator, which automates Slurm cluster management, saving dozens of DevOps hours per week • No buffer capacity charges, as $NBIS maintains a shared spare-node pool, cutting cluster costs by 10-20% Even when using higher-end NVIDIA B200 GPUs, these combined efficiencies result in hundreds of thousands in savings per training run compared to industry-standard providers. 5) $NBIS isn't trying to be another general-purpose cloud, it's building a vertically integrated AI cloud purposely engineered for training, fine-tuning, and inference of foundation models. By owning the full stack (custom hardware, proprietary software, and energy-efficient data centers), $NBIS controls utilization, uptime, and total cost efficiency. Most people don't realize it yet, but the real race in AI infrastructure isn't about who has the most GPUs, it's about who can extract the most value from every GPU hour. All in all, $NBIS is positioning itself as one of the most efficient AI infrastructure providers in the market, tackling the next bottleneck in AI scaling: cost per model trained.

DataOne is expected to unveil a “game-changing” innovation for $NBIS data center in New Jersey soon. We don't know any details yet, but based on a few comments from the company's CEO, I believe it could be related to power consumption. While discussing the future of the industry, he recently said: “We may have other tech invented requiring a lot less power to give the same or better outputs.” For now, I'm just speculating. But if that's true, it would directly address one of the biggest bottlenecks in the data center industry today.

🚨 JUST IN: $NBIS provides financing update. The company plans to raise: - $2B through convertible senior notes - $1B through the sale of Class A shares I've always said we should expect dilution, so this is natural and was already factored into my investment thesis. "Our core AI cloud business is growing rapidly, and we continue to be laser focused on developing this business. We view long-term contracts such as the one we announced this week with Microsoft as important incremental sources of growth. We intend to use the net proceeds from the offerings we have announced today, if consummated, to finance the continuing growth of our business, including the acquisition of additional compute power and hardware, securing strategic high-quality and well-located land plots with reliable providers, the expansion of our data center footprint, and for general corporate purposes. We believe this will enable us to aggressively grow our core business in 2026 and beyond as we aim to scale our global data center portfolio, including through new greenfield sites, and the expansion of our customer base, from AI native tech startups to larger enterprises. As we continue to access the capital that we need to capture the substantial growth opportunity ahead, we will remain focused on mitigating shareholder dilution and maintaining appropriate levels of debt on favorable terms. Our strong balance sheet and our recently announced agreement with Microsoft give us an opportunity to further optimize our capital structure and cost of funding. For example, we expect to finance the associated capital expenditures through a combination of cash flows under the arrangement and the issuance of debt secured against the contract and related infrastructure, at terms that reflect the credit quality of the counterparty. We remain focused on diversifying our funding sources and believe that our June 2025 convertible notes, and the convertible notes we are offering today, enable us to access significant pools of capital. Our convertible notes have been designed to limit potential shareholder dilution through high effective conversion premium levels, and to minimize financing cash costs while preserving flexibility to manage our balance sheet proactively through tailored call features. Each time we have accessed the capital markets, investors have supported our efforts to move at a faster pace to expand our capacity, enabling us to more effectively meet the strong and growing demand for AI compute. We believe our overall financing approach will position us well to accelerate the growth of our core business as we continue to build one of the world’s leading AI infrastructure companies."