AI has arrived and how! It has taken centre stage as the next-gen general purpose technology and is rapidly changing the world as we know it. It is now a force upending markets, business models, capital allocation decisions and even the way we need to plan for our future energy needs. Companies are racing to fold AI into their operations, and some are even learning that adoption without discipline comes with costs.As this shift plays out, the vocabulary of AI is no longer confined to tech blogs and research papers. It is getting mainstream and quietly infiltrating boardrooms, earnings calls and other market literature. Terms like ‘tokenomics’, ‘agentic AI’ and ‘open-source’ are increasingly becoming how companies describe their spending and strategy.So, as investors, it is a good time to get started on these terms and this article aims to help you get warmed up on what they mean and why they matter economically. Read on to find out.TokensAs humans, the smallest unit of information we use to represent language is a single character — which could be a letter or a number or a punctuation mark. But the smallest unit of language that a large language model (LLM) uses is called a token. A token can be a set of characters, or a part of a word or a syllable or a whole word or even a short multi-word phrase. The code that converts language to tokens is called a tokeniser.A good rule of thumb is that an English word is represented by approximately 1.5 tokens. So, if you input a prompt that is 100-word long, the LLM will likely work with 150 tokens. Conversely, one may say a token is about two-thirds to three-fourths of a word. However, this is just to give some perspective. For image-/audio-/video-based inputs, it is not possible to apply this thumb rule. After the input is tokenised, the model then converts the tokens into vectors, which are a long list of numbers. These vectors form the basic computational unit on which the model performs calculations to generate output.That said, let’s understand tokenomics. When it comes to LLMs, model providers such as OpenAI and Anthropic bill customers by tokens, which directly relate to the amount of processing their GPUs perform for a given prompt. Higher the tokens involved, higher the workload on the GPUs which in turn translate to higher cost for the model provider. This billing by tokens is more true of enterprise customers who use APIs (application programming interface) to connect their enterprise software to an LLM than of consumers like you and me.Just as electricity bills are based on units consumed, LLM billing is based on tokens consumed. However, electricity bills change based on what kind of consumer you are. Industrial consumers pay more than what residential consumers do. Similarly, pricing of the LLMs depends on the complexity of the model. A larger, more robust model will cost more per token and vice-versa. Typically, users are billed per million tokens and further, there is a difference based on whether the token is part of the input or part of the output. Input tokens cost lower than output tokens because the amount of processing required to generate a text output, for instance, is higher than reading a text input.Over the past year or so, companies have pushed their workforce to adopt AI in their workflows, often without capping token consumption. Employees were even tracked and ranked based on token usage, as a proxy to measure productivity and to evaluate whether they have become ‘AI-native’. This led to a phenomenon known as ‘tokenmaxxing’. This often incentivised employees to use larger/ reasoning models to boost token consumption even when the problem didn’t warrant their use. Do note that the reasoning steps also consume tokens even though they don’t show up in the output as such.An unnamed enterprise customer of Anthropic was reported to have burned $500-million worth of tokens just because it failed to set spending cap per employee. Similarly, Uber reportedly spent its entire 2026 AI budget in just four months. Its COO even commented that they haven’t yet been able to draw a connection between the company’s use of Claude Code and the amount of consumer features it is shipping. This has led the company to set a cap of $1,500 per employee per month per AI tool in early June. This is becoming an observable trend in other companies too. Ford is reported to have rehired over 300 veteran quality inspectors after investments in AI failed to live up to expectations.While this retreat may turn out to be a short-term fallout of ‘tokenmaxxing’, cost per token appears to have dropped significantly over the last three years — by over 90 per cent, going by some estimates. It remains to be seen whether a balanced approach (after spend caps) towards AI spends results in a favourable cost-benefit equation for enterprises, going forward.Agentic AIAgentic AI is one of the most powerful applications of LLMs. It refers to the ability of an AI model to spawn and deploy agents meant to carry out a complex task. Here’s how an agent is defined. Agents are autonomous systems that can operate independently, use various tools to accomplish complex tasks, planned and executed step after step using the reasoning abilities of an LLM.Here’s an example. Let’s say a user prompts an LLM with agentic abilities ‘Create a to-do list app’. Here the model, without depending on the user (autonomously), internally creates a plan of action, enabled by its ability to reason. The plan could look something like:* Define the features of a to-do list app* Write the code* Test the code* Debug the code* Test and debug the code until the app is stable* Deploy the appThe model will then execute all milestones on its own, turning to the user only if it needs additional inputs. This is a simple way to understand how a typical coding agent works.You may wonder that a chatbot, too, can pull this off. But it may need a detailed prompt from the user that basically incorporates the above milestones and guides the model what to do in a sequence. It may not independently carry out the task in the absence of such a prompt. Autonomy, ability to plan and reason, ability to handle complex tasks – all are defining characteristics of an agent.Another key feature of an agent is the ability to call the right tool to discharge the task. Let’s say a user prompts a model to find out how much an Indigo flight ticket to Goa costs on a Saturday when he is free in the immediate future. The model has to first know when he is free. It needs access to his calendar app (the tool) to know the same. Next, it will write the code to call his Google Calendar using something called an API. An API is simply a piece of code that lets the model talk to a tool. It then runs a code until it finds out the very next Saturday the user doesn’t have any appointments on. Knowing this information, it then connects to the API of another tool — this time to that of a travel booking website (say, Skyscanner) to collect the ticket prices of Indigo flights to Goa on that specific Saturday. Once done, returns the findings to the user — all done without much intervention from the user.While agents are powerful, here’s a key factor one has to remember. When an agent reasons, plans, instructs itself the steps or runs iterations of the code until the task is done — all these consume tokens, even though they don’t show up in the output. Also, when the agent goes through multiple iterations, it will essentially try to keep the earlier iterations also in memory and such iterations become input for the next iteration. Thus, the tokens consumed in earlier iterations keep adding up after each iteration. This makes agentic AI token guzzlers when not managed efficiently.Proprietary vs open-source modelsA proprietary LLM is developed and owned by the organisation that develops it. The training code, training data and the model’s parameters (which determines its size – whether it is a large or small model; parameters are often used interchangeably with ‘weights’) are private and not available for access by outsiders. Popular LLMs such as OpenAI’s GPT series, Anthropic’s Claude Sonnet, Claude Opus series and Google’s Gemini are examples of proprietary models.On the other hand, an open-source model is available for commercial use by persons other than the developer. Basically, one doesn’t need a licence from the developer to use the model. DeepSeek-R1, Alibaba’s Qwen3.6, Meta’s Llama 4 and Mistral 3 are popular examples. Readers must observe a nuance here. In the traditional sense, ‘open-source’ would mean that the source code is accessible. However, in the AI world, often open-source models are simply open-weight models — meaning only the model weights are open-source while the training data and training code remain private. This is true of the models mentioned above. Essentially, the difference between a proprietary model and an open-source one boils down to whether the weights are open-source.What are weights then? Since it could get too technical, here’s what you need to know. Weights (or parameters) determine which piece of input information a model must pay attention to. These are set by the developer during the training phase. Higher the number of weights, larger the model. For context, DeepSeek-R1 reportedly uses 671 billion parameters.Say, a user gives a prompt that reads ‘Einstein proposed the theory of’, a good LLM will predict the next token as ‘relativity’ and deliver the output. This happens because once the LLM sees the tokens ‘Einstein’ and ‘theory’, it will also see the highest weight assigned to the token ‘relativity’. If the model hadn’t been properly trained and a random token, say, ‘gravity’ is seen to be carrying higher weight by the model than for ‘relativity’, it will then hallucinate and deliver ‘gravity’ instead of the accurate ‘relativity’.In an open-weight model, the user can ‘fine-tune’ the weights according to his preferences and build a new product or app with such model. Models such as DeepSeek-R1 let users to even train a new model altogether (typically a smaller model) based on the original model. The model so derived is called a distilled model. For a text-to-speech application strictly for Indian languages, for instance, a user can prune some weights off the larger model (based on which contribute least, thus shrinking its size) and then fine-tune (train) the weights left behind entirely on Indian languages.Reduction in the number of parameters leads to reduction in the model’s size, eventually resulting in the need for lower compute relative to running the original model as such. This particularly makes sense for specialised applications such as drug discovery, flight scheduling for pilots, etc., where the model is not expected to serve other use-cases.Broad differences between using an open-source model as against a proprietary model are captured in the table.Scaling lawsAs far as LLMs are concerned, research has shown that there is a predicted rise in a model’s performance (accuracy of the prediction) when the following three are scaled up – model’s size (parameters), training data and compute.However, there exists a delicate balance among the three and over scaling of any one of them will not yield meaningful improvement in performance. Think of this like a chemical equation. Assume the equation states you need one portion of molecule A, two portions of molecule B and three portions of molecule C to get one portion of compound D. If you want to scale the output to two portions of compound D, you will need twice the amount of all three input molecules — that is, two portions of A, four portions of B and six portions of C. More or less of any one of the three input molecules will not result in compound D, which is the desired result.In its research paper known as the Chinchilla paper, DeepMind researchers found that for compute-optimal training, the model size and the training data should be scaled equally: For every doubling of model size, the number of training tokens should also be doubled. They found that a smaller Chinchilla model with only 70 billion parameters trained on the same amount of compute as some larger models (like Gopher (280 billion parameters) and GPT-3 (175 billion parameters)), outperformed the larger models because it was trained on over four times the training data as the larger models. This means that the larger models had over-scaled on model size before scaling the training data.That said, can the three factors that impact a model’s performance be scaled indefinitely? No, as there will likely be bottlenecks. In the past, there have been bottlenecks. In the 2000s, compute was a bottleneck until researchers recognised GPUs could train models instead of CPUs. In a recent case, limitation to scale models was overcome by the transformer architecture – an innovation in model architecture itself. It was first put forward in 2017 in a seminal research paper by Google researchers called ‘Attention is All You Need’, which now underpins modern LLMs such as Claude, GPT series and Gemini. The architecture basically allows all tokens in a sequence to be processed simultaneously, which earlier architectures couldn’t.Today, AI labs like OpenAI and Anthropic, and hyperscalers such as Amazon, Microsoft and Google are investing over a trillion dollars to scale the compute factor. However, training data could be a possible bottleneck, going forward. Assuming today’s models have trained on all digitally-available data and assuming that incremental data could be contaminated by gen-AI content (gen-AI usage is more prevalent in content created today than before 2022 when ChatGPT was rolled out), scaling quality training data could soon hit a bottleneck. If that materialises, there is a chance that incremental compute being invested on now may not meaningfully yield the desired performance uptick.Artificial general intelligenceAGI or artificial general intelligence is another phrase that readers could stumble upon quite often and all major AI labs of today – OpenAI, Anthropic and Google DeepMind invariably have reaching AGI as their goal. AGI is when an AI system can match or surpass the cognitive abilities of humans. When that happens, AGI can autonomously learn, reason and transfer knowledge seamlessly from one unfamiliar task to another without human intervention. AGI is also referred to as ‘strong AI’.On the other hand, we have ‘weak AI’ or artificial narrow intelligence (ANI). These are systems that are designed to excel at a single, predefined task. A chess bot or a facial recognition software in a security device or a piece of code that automatically classifies an e-mail as spam are good examples. They cannot write essays or generate images or process medical scans.Where do today’s LLMs fall? They are still technically ANI but occupy a middle ground that challenges the traditional definition of ANI. LLMs, today, are versatile. They can do math, generate multi-modal content (not just text, but also images, video), translate and so on. However, companies like IBM still regard them as narrow intelligence and hold that today, AGI is nothing more than a theoretical concept. LLMs may be good at predicting the next most statistically probable word but are far from autonomously achieving the level of spatial reasoning, fine motor coordination, social-emotional awareness and so on that human beings are capable of.Context windowIn the realm of LLMs, the context window is the equivalent of its working memory. It determines how long of a conversation the LLM can carry without forgetting details from an earlier exchange. It is measured in tokens.Let’s look at a simplified illustration. Assume an LLM’s context window is 600 tokens. The user inputs prompt ‘A’, which is 150 tokens long. The model then gives response ‘B’, which also is 150 tokens long. The user adds another exchange ‘C’, which is 150 tokens long. The model responds with an output ‘D’, which is again 150 tokens (things happen very differently in reality, but this simplification helps understand the concept). Now, A, B, C and D fill the model’s context window. The moment the user inputs another 150-token-long prompt ‘E’, it exceeds the model’s context window, leading it to forget or drop the older exchange ‘A’ from its working memory. It may now have to make do with whatever elements of ‘A’ that may have continued into ‘B’ through ‘E’ to work on an output. However, this will, at best, be an educated guess and may lead to hallucinations.Context windows have expanded over time. From something like 2,000 tokens long to over 1,28,000 tokens an IBM Granite 3 model uses, for instance. GPT-5.5 can accommodate over one million tokens in its context window. A million tokens could roughly be about 7,00,000 words and could feel like a lot. For context, the story you are reading right now is 3000 words long. However, today’s LLMs do not just stop with simple text prompts. They can read thousands of lines of code, read hundreds of PDFs which may have images too and also have agentic capabilities. When you consider these, the million tokens may not be a lot.So, are longer context windows the solution? Well, there are constraints. The obvious one being compute. Essentially, as the number of input tokens doubles, it results in the model requiring four times the processing power to handle them. A longer context window also overwhelms the model, leading to performance issues. This is because LLMs are typically said to exhibit a U-shaped attention curve — focusing heavily on the information at the absolute beginning and end of a long prompt, relative to the attention paid to the middle part. This is the same reason, why longer context windows could go against the safety of the model. Malicious actors can take a 50,000-word document and bury a harmful payload or an adversarial system instruction deep in the middle.Published on July 11, 2026