small companies, big models, and the bubble we live in
i’ve been thinking a lot about bubbles lately. the hype cycle, the money chasing the story, the press coverage, the endless articles about what’s next. there’s something intoxicating about it, and there’s something terrifying about it. it’s obvious when you step back and see it from a distance, but when you’re inside, building, shipping, worrying about bugs and deploys, it’s a different view. you feel both insulated and exposed at the same time.
yesterday was a perfect example of that duality. we were trying to upload our latest sage MoE model to huggingface. not for publicity or anything, not for hype, but just so we could test deployment workflows and make things easier for our small internal team. simple enough, right? it turns out the model was too big for a private repository on the free plan. something we should have checked, but it caught us off guard. hours of waiting, trying different configurations, debugging errors, wondering if the platform just hates us or if we messed up. finally, after some frustration and frantic discord (yes, we use discord and not slack at SAGEA) messages, we upgraded to an enterprise plan. a few clicks later, it was live. the relief, the laughter, the ridiculousness of the problem itself. a multi-billion parameter model stranded on a free tier. it’s funny in hindsight, but in the moment, it was panic mixed with absurdity. a small technical hurdle that perfectly mirrors the paradox of building something meaningful while operating on extremely limited infrastructure.
moments like that make me reflect on what it means to be a smaller AI company in 2025. there’s a lot of talk about the AI bubble, and people throw the term around loosely. valuations going crazy. hype around every new model. the narrative that any idea in AI is suddenly worth hundreds of millions. and yet, we’re here. shipping models, building reasoning agents, actually trying to solve problems for users, not just get clicks or press.
being small in this bubble is both challenging and protective. we don’t have millions of dollars to burn on PR campaigns or fancy offices. we don’t have a huge media apparatus spinning stories. we have a model. we have a team. we have execution. the upside is that we are forced to be efficient. every compute allocation, every model iteration, every line of code matters. the downside is that investors, the press, the hype, they often don’t notice. hype is louder than quality. but that’s fine. we’re in this for the work itself. the bubble makes it harder to be noticed, but it also filters out noise in a way. if you survive the early years without burning cash just to signal, you develop resilience. you develop discipline. you develop clarity about what really matters.
i have a raw perspective on this. bubbles are not bad in themselves. they create energy. they create attention. they create the kind of acceleration that can make things happen faster than they would naturally. but they also distort perception. a lot of AI companies right now are built more on perception than product. valuations reflect sentiment more than execution. and that will correct eventually. when it does, the ones who are truly shipping, who have actual models, products, and users, will stand out. and the ones who are chasing headlines will vanish.
for a small company like ours, the MoE research we’re doing is an example. it’s easy to look at larger models with billions or hundreds of billions of parameters and feel small. but the efficiency gains, the architecture choices, the routing strategies we’ve implemented in SAGE MoE make the models outperform some larger counterparts. it’s not hype. it’s reality. it doesn’t generate clicks. it doesn’t create press articles. but it creates leverage. it creates a foundation. when the bubble corrects, or when attention shifts, this is the kind of work that matters. the kind of work that people will actually adopt, integrate, and build on.
the enterprise plan incident also reminded me of something else: infrastructure is both a blessing and a curse. on one hand, platforms like huggingface, colab, and cloud GPU providers democratize AI research. small teams can train, deploy, and distribute models with unprecedented ease. on the other hand, limits on free tiers, quotas, pricing tiers, and access control mean that the tiniest oversight can become a blocker. it’s a reminder that scaling AI is not just a research problem, it’s a systems problem. hardware, software, deployment, team coordination, logistics. all of it matters. sometimes more than the model itself.
there’s a human aspect too. as a founder, you oscillate between feeling excited, terrified, validated, and frustrated, often all in the same day. yesterday’s HuggingFace upload problem could have been solved in thirty minutes if we had planned differently, but it wasn’t. it took hours of collaborative troubleshooting. frustration turned into laughter. and by the end, there was a shared sense of triumph over something so small, so ridiculous, but so real. these moments are tiny milestones. they remind you that building an AI company is not just about the models. it’s about teamwork, resilience, and humor in the face of absurdity.
this is part of a bigger truth. AI is noisy right now. the press cycles every new model like it’s the second coming of technology. investors chase stories. founders chase metrics that aren’t meaningful. but the underlying reality is that progress is still slow, incremental, messy, and human. the majority of small AI companies are not in the headlines. they’re in labs, in Slack chats, in Colab notebooks, in messy deployments. they are shipping. they are learning. they are surviving.
for us, survival is intertwined with execution. every MoE iteration, every agentic CLI prototype, every benchmark we run, every small user interaction matters. the enterprise plan purchase was absurd, but necessary. it’s a small reminder that the infrastructure reality is unavoidable. AI is not just research. it is engineering, systems, operations, and logistics. all of it matters equally to the final outcome. ignoring any part of it is a trap.
and then there’s perception. the bubble, the hype, the social media narratives. i’ve been watching companies get inflated and deflated, sometimes within months. some are hyped for solving nothing, some are quietly building products that will outlast the noise. the problem with perception-driven growth is that it teaches the wrong lessons. hype feels like success. it creates pressure to always signal, always produce content, always move fast enough to be seen. for small AI companies, resisting that pressure is critical. focus on building. focus on execution. focus on the real metrics that matter. users, performance, reliability. attention is temporary. real work is lasting.
there’s also the absurdity of timing. a few months ago, enterprise plans on HuggingFace felt like a luxury. now, they are a necessity. compute costs feel like a looming tax. small mistakes are expensive in time, not money. every decision about model size, checkpointing, deployment, and architecture is weighted by the constraints of being a small team with limited resources. these constraints are frustrating, but they’re also clarifying. they force you to make tradeoffs, to prioritize. they force you to design better systems, not just bigger models.
people often ask me if being a small company in this bubble is scary. and the answer is yes, but in a good way. it sharpens you. it forces you to be efficient, resilient, and creative. it forces you to develop strategies that larger teams can ignore because they have the luxury of scale. we move fast not because we want attention, but because speed is survival. and speed is compounded when coupled with discipline and execution.
the funny thing is that these struggles are also invisible. the market doesn’t see the enterprise plan purchase, the hours spent debugging deployment scripts, the Slack threads discussing checkpoint sharding. what they see is the model size, the benchmark, the press article. small victories are hidden. the work is real. the output is real. but perception is skewed toward the flashy, the loud, the easily digestible.
this is why personal reflection is important. as a founder, you need to separate your confidence from perception. confidence comes from execution. confidence comes from shipping. confidence comes from solving real problems, even small ones, that compound into a meaningful trajectory. perception is fleeting. it can lift you, or it can crush you. it does neither for those who are shipping quietly.
and then there’s the community layer. small companies often rely on communities more than large ones. the feedback, collaboration, and organic adoption are critical. we’ve had discussions with early adopters about sage MoE routing strategies, training efficiencies, inference speed. their input, questions, and experiments help refine what we build. this feedback loop is vital, and in many ways more important than short-term media attention. it’s a real metric of value.
so what does a small AI company do in a bubble? it focuses on three things. first, execution. ship models, ship products, ship benchmarks. second, efficiency. every line of code, every checkpoint, every deployment decision matters. and third, resilience. survive the hype, survive the mistakes, survive the infrastructure limitations, and keep learning.
these lessons are not just abstract. they are lived. yesterday, it was an enterprise plan purchase. tomorrow, it might be fine-tuning the MoE model for 20 more hours without crashing a GPU. next week, it might be testing agentic prototypes under different workloads to see if reasoning consistency holds. it’s endless, iterative, messy. but also deeply satisfying.
there’s also reflection on scale. smaller teams can pivot faster. decisions are more agile. mistakes are caught faster. execution is visible. but you’re also more exposed. one misstep in infrastructure, one failed benchmark, one miscommunication can be magnified because there’s no buffer. the enterprise plan purchase is trivial in isolation, but it’s emblematic of the balance between scale and agility.
i have to remind myself constantly that attention is not the goal. building is the goal. perception is ephemeral. execution is permanent. and the bubble amplifies this lesson. hype teaches you what doesn’t matter. discipline teaches you what does. for small AI companies like ours, these lessons are more valuable than headlines, more valuable than temporary investment chatter, more valuable than any social media metric.
and the human side cannot be ignored. leadership is managing both technical execution and team morale. yesterday could have been a moment of frustration, but it became a moment of laughter. the team learned, the systems improved, and the work continued. resilience is not just about surviving failure, it’s about turning absurd obstacles into teaching moments, bonding moments, and creative solutions.
the AI bubble will pop, inflate, or morph, as bubbles always do. valuations will fluctuate. media narratives will shift. hype cycles will change. but the companies that survive will be the ones with execution, with discipline, with products and users that actually matter. for small AI companies, the advantage is clarity. the disadvantage is exposure. managing both is the work itself.
so we keep shipping. we keep testing. we keep iterating. enterprise plans, GPU quotas, model checkpoints, MoE routing, agentic prototypes, internal tooling, community engagement, feedback loops. every tiny element compounds into momentum. every small victory, like finally getting a model to upload, is a reminder that progress is tangible, real, and meaningful.
being in this space in the current time is so surreal! the responsibility, the visibility, the pace, the unpredictability. it forces reflection, decision-making, and humility. it forces you to recognize that infrastructure hiccups, miscommunications, and random absurdities are just part of building something real. the bubble is noisy. the market is irrational. the press is loud. but execution, shipping, community, and resilience are what endure. and these are things we can control.
i write this today to record perspective, to process what it feels like to operate a small AI company in this moment, to reflect on absurd moments like yesterday, to share a raw take on the bubble, and to highlight the reality behind the headlines. the MoE work, the deployments, the tiny infrastructure challenges, the human dynamics, the feedback loops; all of it is real. it all matters. it all compounds. and in a noisy, inflated, hype-driven market, that is the kind of signal that survives.
ciao, basab
Want to get notified everytime i write stuff? Sign up to my newsletter