11-19 更新 实时
首页/ 算力/ 正文

Booz Allen转用Nvidia的Nemotron

算力 来源:The Information AI 2026-11-11 05:14 0 阅读 机器翻译
Booz Allen转用Nvidia的Nemotron

NVIDIA 正应对其三十年历史上最严重的声誉打击,起因是有爆料称其 CEO 黄仁勋与一名据报为中国间谍机构工作的女性有婚外情(点击此处阅读我们的最新报道)。但对这家 AI 芯片巨头而言,也有一个亮点:其打造全球最佳开源模型的努力已出现初步进展的迹象。

市值 100 亿美元的政府承包商 Booz Allen Hamilton 的首席技术官 Bill Vass 告诉我,几个月前,该公司已从主要使用 Anthropic 的模型来运行其网络安全和国防软件,转向使用 NVIDIA 的 Nemotron 模型。

他说,这一转变带来了显著的成本节省,不过公司拒绝透露具体细节。(据一位熟悉该公司的人士称,Booz Allen 每年在 AI 上的支出达数千万美元。)

“我们一直在做的,与 tokenmaxxing 恰恰相反,”Vass 在接受采访时说。tokenmaxxing 指的是尽可能多地使用 AI。

Vass 说,在实际编写(而非运行)软件时,Booz Allen 的开发者大多仍使用 Anthropic 的 Claude Code,除非客户要求他们改用其他编程工具。这很可能是因为 Claude Code 仍被誉為软件工程领域最好的工具。不过 Vass 表示,随着 Nemotron 模型能力的提升,他有兴趣让公司的开发者用它们来写代码。

Nemotron 之所以能在承包商和政府机构中不断打开局面,部分要归功于它与 Palantir 的关系。这家软件公司近来与 Nvidia 的合作更加紧密,两家公司都把自身定位为 AI 供应商,帮助企业避免依赖前沿实验室 Anthropic 和 OpenAI。Nvidia 最近签署了一项大单,将在内部使用 Palantir 的软件来管理其供应链。

Palantir 于 6 月底推出一款产品,旨在帮助美国政府机构和企业通过其软件使用 Nvidia 的开源模型。Palantir CEO Alex Karp 今年夏天告诉我,一些政府客户已经从闭源的前沿 AI 模型转向 Nemotron。

尝试 Nemotron 的不只是大客户。开源编程软件平台 Kilo Code 表示,10 月通过其软件处理的约 50 万亿 Token 中,Nvidia 的 Nemotron 模型占了近 17%,高于上月的 8%。个人开发者以及来自中小企业和大企业的开发者,主要使用 Kilo Code 来访问开源模型以完成编程工作。

尽管取得了这些进展,Nvidia 的模型仍被前沿模型所压制。过去一个月,Nemotron 模型仅占 OpenRouter 处理 Token 的约 4%,略低于 8 月某一周 7% 的峰值。OpenRouter 是 Stripe 旗下的软件,可让企业在一个平台上自主切换不同模型,其处理的 Token 量远多于 KiloCode(每月约 700 万亿),且覆盖的任务范围比编程更广。

11 月至今,Anthropic 在 OpenRouter 客户支出中所占份额(31%)高于任何其他 AI 供应商,而 Nvidia 的支出占比不到 1%。

Token 作为福利

这些 OpenRouter 统计数据表明,大多数公司仍然把 AI 预算中的最大份额花在 Anthropic 的模型上。随着这些公司设法控制支出,它们在是否对软件工程师设定公司统一的 Token 上限这一问题上出现了分歧——正如我今天与 Phoebe 和 Catherine 合写的这篇报道中所述。

这正成为工程师们的一个症结。人力资源软件初创公司 Syndio 的 CEO Maria Colacurcio 表示,Syndio 的大型企业客户中,负责人才招聘的负责人已开始收到新职位候选人提出的要求:把更高的 Token 预算作为聘用条件的一部分。这些要求出现在候选人拿到录用通知之后,而该录用通知提供了更高或不限量的 Token 预算,以及使用最新 OpenAI 和 Anthropic 模型的权限。

“于是就有了各种各样附带条款,候选人基本上能拿到对这些特殊安排的认可协议,”她解释道。

目前,这些还不算正式的福利。“没有人把它写下来。它并不是政策的一部分,”她说。

如果 AI 价格持续高企,而企业进一步限制支出——苹果、微软和迪士尼等公司已经这样做——Token 有可能成为一种正式化、令人向往的员工福利。

本周三,一起来解读这 50 家有望定义科技下一个时代的早期初创公司。The Information 的记者将带您深入了解值得关注的公司,以及正在塑造未来科技格局的趋势。

译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(Booz Allen Switches to Nvidia’s Nemotron)

Nvidia is managing the worst hit to its reputation in its three-decade history from the revelations that its CEO Jensen Huang had an affair with a woman reportedly working for China’s spy agencies (read our latest on that here). But in one bright spot for the AI chip giant, there are early signs of headway in its efforts to build the world’s best open-source models.

Booz Allen Hamilton—the $10 billion-market capitalization government contractor—several months ago switched from primarily using Anthropic’s models to run its software for cybersecurity and defense to using Nvidia’s Nemotron models, CTO Bill Vass told me. 

That move has resulted in significant cost savings, he said, though the company declined to share specifics. (Booz Allen spends tens of millions of dollars annually on AI, according to a person familiar with the company.)

“What we've been doing is the opposite of tokenmaxxing,” the term for using as much AI as possible, Vass said in an interview.

When it comes to actually building, rather than running, its software, Booz Allen developers still mostly use Anthropic’s Claude Code unless its customers ask them to use other coding tools, Vass said. That’s likely because Claude Code is still lauded as the best tool for software engineering. However, Vass said he is interested in having the firm’s developers use Nemotron models for coding as their capabilities improve.

Nemotron has been making inroads with contractors and government agencies in part thanks to its relationship with Palantir. The software firm and Nvidia have partnered more closely lately, as both companies have positioned themselves as suppliers of AI that can help enterprises avoid relying on frontier labs Anthropic and OpenAI. Nvidia recently signed a big deal to use Palantir’s software internally to manage its supply chain. 

And Palantir in late June launched a product aimed at helping U.S. government agencies and enterprises use Nvidia open source models through its software. Some government customers have already switched from closed-source frontier AI models to Nemotron, Palantir CEO Alex Karp told me this summer. 

It’s not just big customers trying Nemotron. Open-source coding software platform Kilo Code says Nvidia’s Nemotron models accounted for nearly 17% of the roughly 50 trillion tokens processed through its software in October, up from 8% the previous month. Individual developers as well as those from small and large enterprises primarily use Kilo Code to access open source models for coding work. 

Despite that progress, Nvidia’s models are still eclipsed by frontier models. Nemotron models accounted for just about 4% of tokens processed through OpenRouter over the past month, down slightly from a peak of 7% one week in August. OpenRouter, the Stripe-owned software that lets businesses autonomously switch between different models from one place, processes a lot more tokens than KiloCode (about 700 trillion monthly) and for a broader set of tasks than coding.

So far in November, Anthropic accounted for a higher share (31%) of OpenRouter customers’ spending than any other AI provider, while Nvidia accounted for less than 1% of spending.

Tokens as Benefits

Those OpenRouter statistics show that most companies are still spending the lion’s share of their AI budgets on Anthropic’s models. As those firms look to moderate spending, they’re diverging on whether to dictate firm token caps to their software engineers, as I laid out in this story today with Phoebe and Catherine.

That’s becoming a sticking point for engineers. Maria Colacurcio, the CEO of human resources software startup Syndio, said talent acquisition leaders at Syndio’s large enterprise clients have begun receiving requests from new job candidates to get higher token budgets as part of their employment packages. The requests come after the candidates are offered the job for higher or unlimited token budgets and access to the latest OpenAI and Anthropic models.

“So there's all of these sort of side addendums being created where candidates are getting basically agreements to these carve outs,” she explained. 

Right now, these aren’t formal benefits. “Nobody‘s writing it down. It’s not really part of policy,” she said. 

If AI prices stay high and enterprises further restrict spending—as companies such as Apple, Microsoft, and Disney have—tokens could potentially become a formalized, coveted employee benefit.

New From Our Reporters

Upcoming Events

Wednesday, November 11 — The Information’s 50 Most Promising Startups

This Wednesday, unpack the 50 early-stage startups positioned to define tech’s next era. Get an insider look as The Information’s reporters reveal the companies to watch and the trends shaping tomorrow’s tech landscape.

More details

正文保留英文原文,标题与摘要由机器翻译。

本文由采集器自动同步

原文发布于 The Information AI(Booz Allen Switches to Nvidia’s Nemotron),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗