Archived

Annapurna Labs was acquired by Amazon — this page is preserved as an archive. Visit fobi.com for current frontier-tech coverage.

Annapurna Labs

Acquired by Amazon · $350M

Annapurna Labs, owned by Amazon, designs Graviton CPUs, Trainium training chips, and Inferentia inference accelerators for AWS. Acquired by Amazon in 2015 for ~$350M; Graviton4, Trainium2, and Inferentia2 shipping. Critical to AWS cost and margin strategy.

Annapurna enables Amazon's vertical integration in AI silicon, designing custom chips that power AWS infrastructure at scale.

annapurnalabs.com Networking & Custom SiliconUSAacquired · AMZN

At a glance

Total raised

Undisclosed

no public funding on record

Headcount

—

no public estimate on record

Headquarters

Austin

USA

Founded

2011

15 years active

Status

Acquired

Ticker · AMZN

Latest valuation

$350M

Acquired · 2015

For subscribers

Subscribe to see the full profile

Including Funding rounds, Bull / Bear thesis, Stock + earnings, Roster changes, Patents, News, and Open roles.

Founding

50% off

/month

91 of 100 spots left

Full

$10

/month

Available once all 100 Founding Member spots are claimed.

Get full access

Already subscribed? Sign in →

Funding history

Total raised

—

Rounds tracked

Latest stage

Acquired

Latest valuation

$350M

Date	Stage	Amount	Valuation	Lead investors
Jan 22, 2015	Acquired	$350M	$350M	Amazon

Stock performance

How they make money

Other (specify in prose)Pattern

Annapurna Labs operates as a captive internal chip design arm within Amazon, not as a revenue-generating business with external customers. It does not sell its silicon directly; instead, its Graviton CPUs, Trainium training chips, and Inferentia inference accelerators are deployed exclusively inside AWS data centers to power cloud compute, machine learning training, and inference workloads. The economic rationale is wholly about cost and margin structure — Amazon designs custom silicon to reduce its per-unit compute cost compared to purchasing off-the-shelf chips from Intel, AMD, or Nvidia. Revenue is not recognized at the Annapurna level. The financial benefit flows through AWS's consumption-based infrastructure model, where customers pay per hour or per second for EC2 instances, SageMaker compute, or Bedrock inference. By eliminating merchant chip margins and optimizing the silicon specifically for AWS workloads, Amazon captures a wider spread between what customers pay and what it costs to deliver the compute. The model scales with AWS utilization — higher fleet-wide occupancy lowers the effective cost per chip. Annapurna's value is measured in margin improvement across the entire AWS compute portfolio, not in direct product revenue.

Generated by deepseek-v4-flash · June 2, 2026Analytical summary · not investment advice

Founder backgrounds

Nafea Bshara

Co-founder; AWS VP & Distinguished Engineer

Holds bachelor's and master's degrees in computer engineering from Technion — Israel Institute of Technology, where he was accepted to the elite Rothschild excellence program. Began his career at Israeli networking chip firm Galileo, which was acquired by Marvell Technology in 2000 for $2.7 billion; he subsequently served as VP of Technology at Marvell. Co-founded Annapurna Labs in Israel in 2011 with Hrvoje Bilic. Amazon acquired Annapurna in January 2015, and Bshara's team went on to build five generations of the AWS Nitro system, three generations of Arm-based Graviton processors, and AWS Trainium and Inferentia ML chips.

Hrvoje Bilic

Co-founder

Met Nafea Bshara at Technion's Rothschild excellence program in Israel, having emigrated from Bosnia. The two subsequently worked together at Israeli networking chip firm Galileo, which Marvell acquired in 2000, and continued to work in networking silicon before co-founding Annapurna Labs in 2011 in Israel. Annapurna's Arm-based SoC platform — designed for networking storage and gateway applications — became foundational to AWS's in-house silicon strategy after Amazon's $350 million acquisition in January 2015.

AI bull / bear

Bull case

Annapurna Labs sits at the center of Amazon's most strategic cost lever: escaping Intel/AMD and NVIDIA margins. AWS now runs Graviton4 across its general-purpose compute fleet, saving 30-40% on instance costs versus x86 equivalents while delivering competitive performance. That margin improvement flows directly to Amazon's bottom line at cloud scale. On AI, Trainium2 and Inferentia2 give AWS a credible alternative to NVIDIA H100s for training and inference workloads — critical as AI becomes the fastest-growing segment in cloud. Amazon is undercutting GPU cloud pricing by 40-50% on Inferentia-based instances, winning cost-sensitive inference workloads from startups and enterprise customers alike. The $350M acquisition price is arguably the best M&A in semiconductor history: Annapurna now drives billions in annual savings and enables AWS to offer differentiated silicon that keeps customers locked into the ecosystem. As AWS pushes further into frontier AI models and custom training clusters, Annapurna's roadmap — not NVIDIA's — determines how fast and cheap Amazon can scale.

Bear case

Annapurna chips remain behind the competitive frontier on raw performance. Graviton4 trails AMD's latest Epyc on single-threaded workloads, limiting adoption in HPC and latency-sensitive enterprise apps. Trainium2 has yet to prove it can train frontier-scale models efficiently — AWS's own Titan model efforts were reportedly underwhelming, and key AI labs like OpenAI and Anthropic still train primarily on NVIDIA hardware. The software ecosystem is Annapurna's Achilles heel: PyTorch and JAX optimizations for Trainium/Inferentia lag behind CUDA's maturity, creating migration friction that offsets any price advantage. Meanwhile, competitors are moving fast — Microsoft is co-developing Maia with OpenAI, Google has TPUv5, and AMD's MI300 series is capturing second-source GPU spend. Annapurna's value is entirely captive to AWS's ability to win cloud share. If AWS loses AI workloads to Azure or GCP, the massive internal investment in custom silicon becomes a stranded cost thesis, not a moat.

Generated by deepseek-v4-flash · June 2, 2026Refreshes quarterly · see disclosures

In the news

Annapurna Labs · last 30 days

No articles ingested yet for Annapurna Labs. Once the hourly news pipeline is live, every article the classifier tags as mentioning this company appears here with its one-line AI summary and sentiment.

Leadership

Nafea Bshara

Co-founder; AWS VP & Distinguished Engineer

Hrvoje Bilic

Co-founder

—

Products

AWS Graviton
A family of Arm-based, server-class CPUs (Graviton4 being the latest) designed by Annapurna for Amazon Web Services and offered across EC2. Graviton delivers strong price-performance and energy efficiency versus x86, and has become a default platform for a large share of AWS compute—including pre-training workloads for foundation-model companies. It is the cornerstone of Amazon's vertical-integration and cost strategy in the cloud.
AWS Trainium
Custom machine-learning training accelerators (Trainium2 in volume, with successors on the roadmap) built to train large models on AWS at lower cost than merchant GPUs. Deployed in large 'UltraCluster' configurations with the Neuron SDK and high-speed interconnect, Trainium anchors Amazon's bid to reduce dependence on Nvidia for frontier-model training and powers major customer and Anthropic workloads.
AWS Inferentia
Purpose-built inference accelerators (Inferentia2) optimized for high-throughput, low-cost deployment of deep-learning models—especially large language and generative models—on AWS. Paired with the Neuron compiler and runtime, Inferentia lets AWS customers serve models more economically than on GPUs for many workloads, complementing Trainium to give Amazon a full custom-silicon stack for the AI lifecycle.
AWS Nitro System
A combination of dedicated Nitro cards, a security chip, and a lightweight hypervisor that offloads networking, storage, and security from the main server CPU, underpinning the performance, isolation, and security of modern EC2 instances. Originating from Annapurna's silicon, Nitro is foundational infrastructure across AWS, enabling bare-metal performance and the confidential, multi-tenant architecture of Amazon's cloud.