Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
The headline is overstated. In January 2025, former Intel CEO Pat Gelsinger said his startup, Gloo, had decided not to adopt and pay for OpenAI’s o1 model for one planned product after its engineers began testing DeepSeek-R1. Gloo’s stated plan was to rebuild that product around an open-model foundation—not simply to switch to DeepSeek’s hosted API, and not to ban OpenAI across the company.
What Gelsinger actually said
Gelsinger’s comments, reported by TechCrunch on January 27, 2025, concerned Gloo’s planned AI service, Kallm. He said Gloo engineers were already running DeepSeek-R1 and that the company had decided not to adopt and pay for OpenAI’s o1 model for the product.
He described a plan to rebuild Kallm “from scratch” using Gloo’s own open-source foundational model. That was a reported intention, not independent confirmation that the rebuild was completed. The account also does not establish that Gloo stopped using OpenAI for every other purpose. Gelsinger was then chairman of Gloo, a messaging and engagement platform for churches; he was no longer Intel’s CEO, having left the role in December 2024 after about four years.
Free tools Windows power users keep installed
One-click scans. No signup required.
So “done with OpenAI” turns a product-level procurement and architecture decision into a much broader personal or company-wide rejection that the evidence does not support.
#1 Best Overall
Why DeepSeek-R1 changed the calculation
DeepSeek, a Chinese AI company, released R1 in January 2025 as a reasoning-focused model. Such models spend additional computation working through a problem before answering. DeepSeek released model weights as well as an API, and also offered smaller distilled versions. Its release notes and research paper presented R1 as comparable to OpenAI’s o1 on selected reasoning tasks.
That combination mattered to companies evaluating AI products: reported strong performance on certain tests, lower launch-era API prices, and the possibility of running or adapting weights rather than relying only on a closed hosted service. The full R1 model was reported at about 671 billion parameters, requiring substantial hardware. Smaller distilled models, ranging from roughly 1.5 billion to 70 billion parameters, offered more practical options, though hardware needs and capability vary by size.
DeepSeek’s benchmark claims should be read narrowly. The company reported that R1 matched or exceeded o1 on selected benchmarks, including AIME, MATH-500 and SWE-bench Verified. Those results do not prove it is better for every task, or establish production reliability, latency, tool use, safety, long-context performance or enterprise support. A benchmark is evidence about the tested setup and task—not a universal ranking of models.
Recommended Free Tools
The price and training-cost claims need context
Contemporary coverage described R1’s API as roughly 90% to 95% cheaper than OpenAI o1 at launch. That is a historical comparison, not a statement about current prices. Actual costs depend on input and output mix, caching, reasoning-token usage, volume and service terms. Self-hosting introduces different expenses: accelerators, power, storage, networking, engineering, monitoring and maintenance. The cheapest token price is not necessarily the lowest total cost for a particular business.
Gelsinger also estimated that DeepSeek’s training was 10 to 50 times cheaper than OpenAI o1’s. That was his assessment, not a verified like-for-like accounting comparison. The widely repeated figure of about $5.5 million referred to a specified DeepSeek training run under particular conditions. It does not represent the total cost of building DeepSeek, its research program, data and infrastructure, or every model in its product family. It would be misleading to conclude that DeepSeek built an equivalent frontier-AI company for $5.5 million.
Gelsinger’s broader argument was that more efficient computation could expand AI use rather than simply reduce revenue for established providers. He pointed to cheaper computing, ingenuity under constraints and open ecosystems as forces that could accelerate adoption. He also argued that capable models could eventually reach more devices, including phones, vehicles, wearables and hearing aids. Those are his strategic interpretations, not outcomes established by R1’s release.
Open weights are a choice with responsibilities
“Open source” is often used loosely in AI coverage. In this case, it is more precise to say R1 weights were released under an MIT license, while distinguishing those weights from DeepSeek’s hosted service and its terms. Access to weights can give an organization more control over deployment, customization and data locality if it runs the model in its own environment. It can also reduce dependence on a single API provider.
But downloadable weights do not make a model effortless or risk-free. The customer must arrange hosting, security, updates, monitoring, abuse prevention, evaluation and compliance review. A full-size model calls for substantial hardware; smaller distilled variants are more accessible but are not interchangeable with the full model. “Free weights” can still produce a large operating bill.
A hosted API, by contrast, can be quicker to integrate and leaves infrastructure operation to the provider, but brings usage fees, provider dependence and questions about data processing and service terms. Gloo’s reported plan points toward an open-model strategy for Kallm; it does not show that the company merely became a DeepSeek API customer.
What the enthusiasm did not settle
DeepSeek’s reported performance and pricing did not resolve questions that matter to buyers:
- Cost and hardware transparency: Some observers questioned whether public figures captured all prior runs, infrastructure and hardware. The cited reporting relayed suspicions about more advanced hardware, but those claims were disputed and not established as fact.
- Privacy and governance: Where prompts, outputs and logs are processed, who can access them, and what contractual protections apply are separate questions from model quality. DeepSeek’s Chinese ownership and data-handling practices were concerns for some Western organizations and should be assessed against each buyer’s requirements.
- Content behavior and compliance: Moderation and political-content behavior may make a model unsuitable for particular markets or uses, even if it is capable and inexpensive.
- Operational reliability: Uptime, rate limits, latency, incident response and service guarantees are distinct from benchmark scores.
- Security and maintenance: Self-hosting gives a buyer more control but also makes it responsible for securing, updating and monitoring the system.
These are evaluation criteria, not proof that a particular service fails them. Buyers should test the model and deployment they would actually use, review current vendor documentation and contracts, and involve security, legal and compliance teams where appropriate.
How a business should make the same decision
For an organization choosing between a hosted proprietary model and open weights, the useful comparison is not “which model won the headline benchmark?” It is whether the model and deployment meet the organization’s requirements at an acceptable total cost.
- Test the real workload. Evaluate representative prompts, edge cases and failure modes, not only public benchmark scores.
- Calculate total cost. Compare API usage with hardware, power, engineering, storage, networking, monitoring and support for the expected workload.
- Choose a deployment model. Compare hosted API, private cloud, on-premises and hybrid options, including where prompts, outputs and logs are processed.
- Review license and terms. Confirm commercial-use, redistribution and fine-tuning permissions, and distinguish model-weight licensing from hosted-service terms.
- Check reliability and safety. Measure latency and availability for the application; test moderation, adversarial inputs and regulated or sensitive workflows.
- Plan for ownership after launch. Decide who handles model updates, security fixes, regressions, monitoring and incident response.
- Account for hardware and organizational capacity. A full-size model and a distilled variant have different infrastructure needs. Open deployment is a poor bargain if the team cannot operate it safely and reliably.
Some buyers will value vendor-managed infrastructure and support enough to choose a closed API. Others will prioritize deployment control, customization or reduced provider dependence and accept the work of self-hosting. There is no universal winner; the answer depends on workload, scale, governance and the team’s ability to run the system.
What the story means—and what it does not
Gelsinger’s move was an early, visible signal that competitive open-weight models could change how companies evaluate AI vendors. DeepSeek-R1 challenged the assumption that only a closed, high-cost service could deliver strong reasoning results. It did not demonstrate that OpenAI was obsolete, that R1 was superior across the board, or that open models are automatically cheaper or safer in production.
The most defensible reading is specific: after testing R1, Gloo chose not to pay for OpenAI o1 for its planned Kallm product and said it intended to build around its own open-model foundation. That makes the story a case study in the changing economics and architecture of AI—not a blanket breakup with OpenAI.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
