Enterprise AI bills are exploding even as the cost per token collapses, creating a paradox that has caught the attention of a new generation of chip makers. While token prices have fallen 98% since late 2022, from $20 per million tokens to just $0.40, companies are consuming tokens at such a rapid pace that their total AI spending has jumped an estimated 320% in the same window. This mismatch between cheaper units and higher bills is reshaping how the industry thinks about AI infrastructure, and it is driving startups to rethink the chips that power inference workloads.

Turiyam.ai, a Bengaluru-based startup founded in 2024, is betting that the next battleground in AI hardware will not be training speed but inference cost control. The company raised $4 million in pre-seed funding in March 2026 to build a full-stack inference platform with custom silicon designed from the ground up for inference workloads, rather than adapted from general-purpose graphics processing units (GPUs) built for training.

Why Are Enterprise AI Bills Skyrocketing Despite Cheaper Tokens?

The disconnect between falling token prices and rising enterprise bills reveals a fundamental shift in how companies are using AI. Per-developer token consumption has grown roughly 18.6 times in just nine months, according to research from Jellyfish, a software analytics firm. This explosion in usage means that even though each token costs far less, the sheer volume of tokens being consumed has overwhelmed any savings from price reductions.

The real-world impact is visible across major tech companies. Uber exhausted its entire 2026 AI coding budget by April. Microsoft revoked developers' access to Anthropic's Claude Code tool months after enabling it, citing cost concerns. One company reportedly ran up a $500 million Claude bill in a single month after forgetting to set usage limits. Priceline employees told TechCrunch that a routine Cursor contract renewal came back four to five times more expensive than the previous year, even though per-token pricing had not increased.

"Customer calls have shifted from 'What can it do? Is it good enough?' to 'We're spending so much. What visibility do you have? What token controls do you have?'" said Alexander Embiricos, OpenAI's head of enterprise.

Alexander Embiricos, Head of Enterprise at OpenAI

This cost crisis has reached the highest levels of the industry. The Linux Foundation announced a new Tokenomics Foundation aimed at standardizing AI token usage and billing, with a formal launch planned for July and new metrics like cost-per-intelligence and tokens-per-watt.

How Does Inference Hardware Differ From Training Hardware?

Most chips on the market today were designed for training large language models and later adapted to handle inference, the process of running a trained model to generate predictions or responses. This dual-purpose design creates inefficiencies. Turiyam's approach is fundamentally different: the company is designing inference chips from scratch, with a hybrid memory design, a compiler-led optimization layer, and a full-stack approach where the chip and software are co-designed together.

The thesis is straightforward: inference hardware built specifically for inference workloads can achieve better throughput and performance-per-watt than a general-purpose GPU doing inference as a secondary task. Because the inference accelerator is optimized for one job instead of two, it can deliver lower total cost of ownership for companies running large-scale inference operations.

"Putting the software stack in place from day one, rather than as an afterthought, is what makes the approach differentiated and relevant for where the market is headed," said Ritu Verma, managing partner at Ankur Capital.

Ritu Verma, Managing Partner at Ankur Capital

Turiyam is not alone in this space. Other inference-focused silicon vendors including Groq, SambaNova, Cerebras, and Fractile have spent the last three years building alternatives to general-purpose GPUs. However, Turiyam's geographic base and capital profile are distinct: the company is an Indian operation building on indigenous compute infrastructure, starting from a much smaller capital base than its global peers.

Steps to Understanding the Inference Chip Market Shift

• The Cost Problem: Enterprise AI spending has grown 320% while per-token costs have dropped 98%, meaning companies are consuming far more tokens despite lower unit prices, creating a structural cost crisis at the inference layer.
• The Hardware Solution: Inference-specific chips designed from scratch can optimize for throughput and power efficiency in ways that general-purpose GPUs cannot, because they are built for one workload instead of two.
• The Software Advantage: Companies that pair custom inference hardware with compiler-led optimization and full-stack software design from day one can deliver lower total cost of ownership than vendors that hand customers a chip and expect them to write their own compiler.
• The Market Validation: Turiyam is testing its thesis through partnerships with NTT Global Data Centers for global deployment and with India's Centre for Development of Advanced Computing for domestic validation on Hindi language models.

Turiyam's $4 million pre-seed round, led by Ankur Capital and Micelio Fund (part of Axilor Ventures), will fund the compiler stack, the first silicon tape-outs, early pilots with select enterprises, and engineering hires needed to prove the architecture on real workloads. The company is currently in the pilot phase with select enterprises.

The startup has already secured two significant deployments to validate its approach. The first is a partnership with Tokyo-headquartered NTT Global Data Centers to host and scale Turiyam's next-generation inference servers inside NTT's global data center facilities, which offer enterprise-grade security and renewable energy integration. This deal gives Turiyam a carrier-grade data center footprint without the capital expenditure of building one.

The second deployment is domestic validation. Turiyam deployed its inference engine on the Rudra 1 and Rudra 2 servers built by the Centre for Development of Advanced Computing (C-DAC), under India's Ministry of Electronics and Information Technology. During validation, a large language model for Hindi covering 37 dialects was executed inside C-DAC's infrastructure, demonstrating that inference-specific hardware can handle complex, multilingual workloads efficiently.

The broader context matters here. Average enterprise AI budgets have exploded from $1.2 million per year in 2024 to $7 million per year in 2026, a nearly sixfold increase in just two years. This growth is not driven by higher per-token costs but by the sheer volume of inference workloads companies are pushing into production. As more use cases move from experimentation to production, the inference layer becomes the critical cost lever, and the chips that run those workloads become the focal point for cost optimization.

For enterprises struggling with runaway AI bills, the emergence of inference-specific hardware represents a potential path forward. Rather than accepting that cheaper tokens automatically mean cheaper AI, companies may soon have the option to deploy inference workloads on chips designed specifically to minimize total cost of ownership, not just per-token pricing.