ML/AI Platform Construct vs Purchase Choice: What Components to Contemplate

ML/AI platforms present the ecosystem for constructing, deploying, and managing the lifecycle of machine-learning fashions and AI providers.

There is no such thing as a one-size-fits-all method to implementing an ML/AI platform:

Constructing an in-house platform will be quick for particular use circumstances and assist maximize return on funding rapidly.

Shopping for a platform or platform elements saves inside growth effort, however it may be a protracted bureaucratic course of and tough to evaluate third-party elements.

Most organizations in the end determine on a buy-and-build answer, the place third-party elements and instruments are built-in right into a customized platform.

An ML/AI platform offers a coherent assortment of instruments and frameworks to construct, deploy, and handle machine studying (ML) fashions and AI providers. Reasonably earlier than later, groups and organizations trying to streamline their ML workflows and convey AI-driven merchandise to market extra rapidly face the query: Ought to they build a custom platform or purchase one?

Given the steep rise in curiosity in AI expertise, the massive cloud firms are investing closely in ML/AI platforms as an integral a part of their product panorama. On the identical time, there’s a variety of specialised instruments and a thriving open-source ecosystem.

The AI landscape is huge and sophisticated, making it tough to navigate the market. Additional, every firm is exclusive by way of the sort of knowledge used, the core enterprise mannequin, compliance requirements, growth processes, deployment infrastructure, and the expertise stack.

Over the previous decade, I’ve helped my groups construct ML/AI platforms from scratch and architected platforms integrating completely different cloud distributors and SaaS options. I’ve participated in numerous vendor assessments and have not too long ago led an initiative to switch an in-house answer with a SaaS instrument.

On this article, I’ll summarize all that I’ve discovered that can assist you determine what to consider when deciding whether or not to purchase or construct an ML/AI platform.

Construct vs purchase: advantages and downsides

Regardless of what advertising can have you imagine, you received’t discover a ready-to-use ML/AI platform off the shelf. In my expertise, you’ll at all times want to mix a number of instruments to reach at an answer that matches your wants.

Thus, “shopping for” an ML/AI platform means buying a number of completely different merchandise, even when considered one of them is an “end-to-end answer.” For the aim of this text, I’m defining “shopping for” as everytime you spend cash buying or licensing a instrument. In any other case, I’ll think about it “constructing.”

When confronted with the “construct vs purchase” determination, at first, the choices can appear easy:

  • Purchase: It’s quick to start out producing worth and cheaper in complete, as no upfront growth is required. Nevertheless, third-party merchandise is perhaps much less versatile in the long term.
  • Construct: Expensive and sluggish to get off the bottom, however an ideal match with the enterprise wants.

Sadly, it’s not that straightforward. Shopping for an ML/AI platform will be sluggish, bureaucratic, and technically demanding to check. Conversely, constructing an ML/AI platform will be fast for skilled groups and supply a great return on funding instantly.

The excellence between “construct” and “purchase” can also be not as clear because it appears at first look: Normally, you have already got an current codebase and inside processes that you should transfer to the brand new platform. And even an end-to-end SaaS answer requires configuration and customization.

The place does open-source software program (OSS) match right here?

Many popular ML/AI tools are open-source, and it’s extremely unlikely that any workforce constructing an ML/AI platform is not going to use some open-source element.

Open-source instruments are interesting as a result of you’ll be able to purchase them with out licensing charges and begin utilizing them instantly. Nevertheless, despite the fact that you don’t pay for OSS software program, it’s nonetheless a third-party element you don’t management. Whereas it’s theoretically attainable to create an in-house fork and totally adapt a instrument to your wants, I’ve not often seen this occur in actuality.

I like to incorporate open-source software program within the “construct” class, because it essentially entails doing your personal integration and upkeep work. Extensively used open-source initiatives usually include documentation and a useful volunteer neighborhood. Nevertheless, there’s often no help accessible past that. Thus, the work required to undertake an OSS instrument into your platform doesn’t differ in some ways from integrating a element you could have written from scratch.

In relation to open-source instruments and platforms (relatively than libraries and frameworks), a lot of them are backed by companies that supply paid help and/or a SaaS model. This lets you delegate the hassle to host and preserve a instrument to a 3rd get together in a while.

You too can think about beginning with a managed service to get operating rapidly and later switching to a self-hosted model in case your necessities exceed the customizability. According to my definitions for “construct” and “purchase” above, I’ll categorize utilizing a SaaS model of an OSS instrument as “purchase.”

Advantages of constructing a customized ML/AI platform

  • Customization: You possibly can tailor the platform to fulfill particular organizational wants and requirements, which is especially related when your knowledge will not be in a normal modality or format. For instance, a number of instruments are biased towards tabular knowledge.
  • Effectivity: You possibly can optimize the platform to streamline growth and deployment. This may contain specialised algorithms, particular pipelines, or explicit necessities for mannequin deployment.
  • Integration: You possibly can seamlessly combine with current programs and workflows because you management the APIs and adapters. This may be essential if your organization upholds excessive requirements of safety and compliance.
  • Governance and compliance: You possibly can implement governance insurance policies and guarantee compliance. Necessities can come from knowledge privateness rules (GDPR and HIPAA) and safety requirements like SOC2, in addition to from inside your group.

Advantages of shopping for an ML/AI platform

  • Value: Shopping for an ML/AI platform eliminates the necessity for upfront investments in growth work and infrastructure. It additionally reduces ongoing upkeep and administration, which eat vital sources and infrequently require new personnel.
  • Time: Shopping for or licensing an ML/AI platform will be fast. You possibly can acquire a critical aggressive benefit by shortening the time-to-market of your AI providers.
  • Experience and help: Licensing plans for ML/AI platforms usually embrace entry to devoted help and specialists who assist with implementation, troubleshooting, and updates.
  • Give attention to core competencies: By buying an ML/AI platform, you’ll be able to focus your sources in your core enterprise. That is very true for organizations the place AI expertise will not be the product however a method to an finish.

Drawbacks of constructing a customized ML/AI platform

  • Complexity: Growing and sustaining a platform is complicated and resource-intensive. It requires an in-house workforce of specialists from a number of areas, akin to knowledge engineering, ML engineering, safety, procurement, and operations.
  • Value: Designing, creating, and deploying a customized platform is a heavy funding. The prices of internet hosting, sustaining, and evolving a platform not often stay fixed however have a tendency to extend over time.
  • Efficiency, scalability, and flexibility: Designing and creating high-scale options that may develop with your online business could be very complicated. Even when your workforce has prior expertise constructing ML/AI platforms, it’s straightforward to underestimate the vary of experience and energy required.
  • Threat of overengineering and scope creep: It’s straightforward to include pointless options or complexity that delivers little or no worth. Including increasingly options can delay the supply and sluggish future growth, lowering utility and growing prices.
  • Safety and privateness: When constructing an inside platform, you could take care of all of the complexity of this subject your self. You need to perceive and implement completely different requirements imposed by clients, domain-specific rules, and the legislation.

Drawbacks of shopping for an ML/AI platform

  • Vendor lock-in: Shopping for an ML/AI platform makes you depending on that individual vendor, restricts flexibility, and impedes supply velocity. Migrating off a core element of your platform in a while will be extremely complicated, particularly when you can’t afford manufacturing downtime.
  • Customization and integration limitations: Many industrial ML/AI platforms supply an in depth vary of built-in options. But, they could not align together with your wants, and their constraints would require compromises. Additional, working a platform the place you solely use a fraction of the options can in the end be worse than working across the lack of 1 particular functionality.
  • Information privateness and safety dangers: Entrusting knowledge to a third-party vendor carries inherent knowledge privateness and security dangers. Imposing your personal requirements can show difficult or downright unattainable. A typical approach out is putting in a SaaS answer in your personal cloud infrastructure, which provides you extra flexibility to fulfill your wants however brings with it most of the drawbacks of the “construct” choice.
  • Restricted management over expertise and roadmap: You received’t have management over a vendor’s underlying expertise stack, roadmap, or priorities. This may affect your group’s future strategic selections and innovation.

Components to contemplate within the construct vs purchase decision-making course of

Investing in an ML/AI platform is a serious determination for any firm. Whether or not you construct or purchase it, the choice entails a multi-year dedication.

Due to this fact, it’s essential to grasp all elements to make an knowledgeable and sound determination. Within the following, I’ll information you thru key areas to analyze and acquire readability on.

Technical experience and organizational maturity

Constructing software program is at all times difficult. This platitude is especially true within the case of an ML/AI platform. Past managing a code base, you’ll additionally have to handle the info layer and fashions.

Whereas conventional software program solely modifications conduct when code modifications, machine-learning fashions behave in a different way when enter knowledge modifications. Additionally, you will probably be working with GPUs and large container pictures. The stochastic nature of ML fashions requires new approaches to testing, monitoring, and observability.  

You may also end up in a company with proficient knowledge science groups that may deal with the intricacies I simply outlined however battle with the fundamentals of software program, infrastructure, and knowledge engineering. That mentioned, constructing software program is usually a web optimistic for small firms, startups, or new groups, because it presents a possibility to stage up on transferable expertise and set up workflows.

Listed below are some inquiries to ask when exploring this space:

  • Organizational maturity: How far are you in your ML/AI journey? Do you could have a transparent roadmap of what options and capabilities you’ll require in a yr? 5 years? The following decade?
  • Expertise and expertise: Do you could have ample software program and infrastructure engineering experience inside your workforce? Do you should rent for added expertise, and the way lengthy will it take? Do you could have established DevOps practices? Is there an current workforce that would deal with platform operations?
  • Non-functional necessities: Do you perceive what it takes to combine or construct a platform in your present infrastructure? What about management administration, versioning, tenant isolation, and scalability? Do you could have somebody who can assess, design, and/or implement it?
  • Investing in experience: Which specialties do you should rent for? Is it lifelike to presently discover these expertise in your labor market? Is it value coaching your present workforce on a specific ability? Will it repay for the group to amass competency in a sure space?


The prices of shopping for or constructing and sustaining an ML/AI platform are definitely first-of-mind for a lot of decision-makers.

Constructing software program from scratch is commonly costly, as we’ve got to consider infrastructure, buying or licensing elements, and salaries.

Infrastructure prices sometimes are the continuing prices for cloud sources, which will be lowered by means of long-term contracts. Some organizations select to put money into on-premise infrastructure or place their very own {hardware} right into a third-party knowledge heart.

When assessing the prices of shopping for a element or instrument, it’s essential to recollect simply how costly the salaries of an inside workforce to construct and run the platform could be. Firms that haven’t operated a large-scale platform are sometimes stunned to find what it prices to have folks on name 24/7 to satisfy service-level agreements (SLAs).

You also needs to think about hiring occasions, alternative prices related to a later platform go-live, upkeep efforts, and getting the workforce on top of things on a brand new expertise stack.

Listed below are some value classes and questions you need to think about:

  • Salaries and consulting charges: Do you could have a forecast of your hiring wants? Do you want extra senior or junior of us? Do you should rent regionally or globally? What salaries do it’s a must to price range for? Will you should work with recruiters? Are you able to think about a contract-to-hire method to attenuate dangers and velocity up growth?
  • Licensing and subscription charges: How will the price change as extra customers are onboarded to a instrument and the variety of managed fashions will increase? Will extra income offset the price enhance of including extra clients to the platform?
  • Storage capability and value: What knowledge do you should retailer as a part of your platform (e.g., efficiency metrics, experiment metadata, datasets, mannequin recordsdata, epoch snapshots, containers)? Do you should retailer all that data perpetually, or are you able to implement retention insurance policies? How a lot knowledge will you retailer per day, month, or yr going ahead? Are you able to make use of chilly storage?
  • Technical debt: Are you clear the place and when you’ll tackle technical debt, significantly when constructing your personal answer? Do you could have a option to observe, worth, and prioritize tech debt? Have you ever factored within the prices of doubtless re-architecting the platform on account of unexpected necessities?
  • CI/CD and growth instruments: Do you could have a dependable estimate of the CI/CD prices? For instance, what is going to it value to often verify containers for vulnerabilities and run integration checks on every PR? How costly will dev instruments, GPUs, and container or bundle registries be?
  • Infrastructure value: Are you aware what infrastructure you’ll want? Do you could have expertise monitoring cloud prices? Will your cloud or platform supplier be capable to provision sufficient sources? Does the SaaS product you’re wanting into permit you to buy extra capability?

Time to market

There is a chance value related to spending time constructing your personal ML/AI platform. Whereas doing it, you aren’t allocating inside sources to areas that present a extra fast aggressive benefit. Constructing an in-house answer can detract from what’s important on your group.

Conversely, buying software program or licensing a platform will be time-consuming. Zoning in on attainable choices, completely assessing them, and negotiating phrases can stretch over a number of months. In case your group is topic to compliance or regulatory necessities, vetting third-party instruments can simply take as much as a yr.

Whether or not you’re constructing or shopping for the ML/AI platform, you should think about the period of the recruitment course of and onboarding time. Hiring AI expertise tends to be a sluggish course of in today’s volatile labor market. You’ll additionally probably want a major period of time to get your current workforce on top of things on the brand new platform or tech stack.

Listed below are some inquiries to get you began desirous about time-to-market:

  • Accessible workforce capability: Do you could have sufficient expertise internally (amount and high quality)? Are you able to rent rapidly sufficient to fill present ability gaps and enhance your workforce’s capability? The place are you able to herald exterior consultants or contractors to reinforce your workforce within the brief time period?
  • Necessities: Do you want a managed answer, or does your workforce favor to put in and function the product? Which ​​non-functional necessities are important on your use case? Do you could have an outlined baseline with which to match a brand new answer?
  • Procurement course of: Do you could have folks in your group who know the right way to conduct a vendor evaluation? Which stakeholders need to be concerned, and when? Are you able to begin utilizing an OSS model of a instrument whilst you’re negotiating a SaaS contract?
  • Improvement course of: Who will lead and steer the platform growth? Does your group have a confirmed observe report of delivering complicated software program initiatives on time? Is your workforce working in a longtime growth course of, or do you should arrange growth groups from scratch?

Upkeep and customization wants

Upkeep and operations are vital useful resource drains, particularly when constructing your personal ML/AI platform.

You need to preserve a roadmap and plan for brand spanking new options. You additionally want to supply help for current functionalities, which in some circumstances requires a devoted workforce. Since there isn’t any third get together or neighborhood round a customized answer, you’ll want to put in writing and preserve all documentation internally.

Whereas specializing in a build-or-buy determination, it’s straightforward to overlook to plan to help the platform’s finish customers. Neglecting this side can rapidly develop into a supply of friction or result in a platform not being utilized correctly.

One other consideration is the extent of customization required to fulfill organizational necessities. A customized answer presents higher flexibility on this regard, whereas shopping for software program might supply built-in options with restricted customization choices.

Listed below are some inquiries to get you began:

  • Customization: What customizations do you want for a given instrument? Are the customizations attainable as a part of the usual providing, or do you should negotiate an addition to the contract? Are you able to implement customizations in-house, or do you should pay the seller or a third-party advisor?
  • Platform product administration: Who will steer platform product growth in the long term? What’s the course of to put, consider, and determine on function requests? How will you’re employed with a vendor to maintain up with product updates and guarantee you aren’t stunned by breaking modifications?
  • Assist for finish customers: How will you onboard knowledge scientists and ML engineers to the platform? Who’s answerable for internet hosting workshops or sustaining academic materials? How will you solicit suggestions? Do you should embed platform specialists with different groups?

Vendor repute and goal market

Understanding a vendor’s enterprise and outlook is essential. Small distributors might have much less secure roadmaps and are likelier to vanish or pivot. Nevertheless, they is perhaps extra prepared to work with you on particular wants. Massive distributors supply extra long-term reliability however usually present generic options that want intensive customization. They will also be sluggish to ship new options if a product will not be their flagship providing.

Assessing a instrument’s or platform’s buyer base, ecosystem, and neighborhood is equally very important. Usually, it’s preferable when you’re a vendor’s goal buyer or the core viewers of an OSS mission. This ensures the seller is ready to work with companies of your dimension and maturity. It additionally makes it extra probably that future modifications to a product will fall consistent with your wants.

I like to recommend discovering a steadiness and shifting ahead with that. It’s unlikely you’ll find an ideal match on your use case.

  • Lengthy-term perspective: What does the product’s roadmap seem like? Are you able to ensure that the instrument or platform will probably be accessible for the years to come back? Will the platform develop together with your wants and future necessities? What function does the platform you’re on the lookout for play in a vendor’s portfolio?
  • Different customers: Who else is utilizing the platform or instrument? Are they corresponding to your group and workforce? How massive is the person base of the OSS element you’re contemplating? What are customers saying in online communities or different on-line locations? Are there present or former customers in your private community? Are you able to attend a conference or a meetup to get first-hand insights?
  • Open-source mission governance: Is the OSS mission backed by a enterprise or basis? Does the mission relaxation on the shoulders of a sufficiently massive variety of maintainers? Are safety points and bugs appropriately dealt with? What’s the course of for suggesting or contributing new options?

Compliance, privateness, and governance

In my expertise, extra technically-minded of us usually underestimate this issue. Nevertheless, implementing compliance necessities can drain your whole supply capability if not deliberate for upfront.

It often requires a company-wide effort to realize compliance with requirements and rules. If achieved correctly, it may open new alternatives and even whole markets as clients search presents that meet their necessities.

Compliance entails following good software program growth practices inside your workforce and platform. Sometimes, particular requirements and processes are built-in into the platform to make sure enforcement. Many of those requirements adhere to security by design principles, akin to minimizing assault floor space, the precept of least privilege, and the separation of considerations.

In case you are topic to strict rules, there’s probably no approach round internet hosting the platform by yourself (cloud) infrastructure. This can enable you full management over your knowledge and implement any extra requirements.

Listed below are some questions you need to know the reply to earlier than deciding on an ML/AI platform:

  • Rules: Are you topic to rules like FedRAMP, SOC 2, HIPAA, or GDPR? What actual necessities do you should fulfill on a technical and organizational stage?
  • On-premise deployment: Are you restricted to self-hosting the platform? Can a platform be put in by yourself cloud infrastructure, e.g., in a devoted VPC? Does a third-party platform help customized registries for containers and packages?
  • Information privateness and governance: What calls for are imposed by rules or clients? Are you allowed to retailer your buyer’s knowledge on infrastructure hosted by a 3rd get together? Do you should management which knowledge can be utilized for which fashions? Do you should observe how knowledge is utilized in fashions and downstream?
  • Software safety: Do you want capabilities to take care of mannequin hallucinations, immediate injection, or malicious person enter? What measures and processes do you should implement to forestall, detect, and mitigate safety points? What ensures can a vendor or OSS mission give?
  • Technical measures: Do you should implement safety requirements on the CI stage? Do you should scan your container pictures often? Do you should have inside base pictures?

My advice

In my expertise, no matter your preliminary method, you’ll converge to a buy-and-build answer. Normally, organizations on the lookout for an ML/AI platform have already got an current codebase and established inside workflows. Thus, you’ll at all times have to combine with the present organizational processes and infrastructure.

Adapting and customizing an ML/AI platform to fulfill inside necessities is at all times mandatory and will probably be an ongoing course of, irrespective of when you purchase a platform or construct it from scratch. The necessity for these modifications sometimes comes from particular knowledge ingestion or deployment necessities.

Value financial savings will quickly develop into a prime precedence. As your group expands, you’ll really feel strain from administration to maintain your cloud invoice and salaries in verify. On this scenario, it pays off in case your preliminary value estimates had been well-founded and you’ll clearly exhibit the platform’s worth to the enterprise.

If I needed to summarize all the pieces I’ve discovered over time right into a single sentence, my advice is to start out as small as attainable and scale up as wanted. Regardless of one of the best efforts, your preliminary price range estimates and understanding of the necessities are incomplete. 

If you happen to rapidly get a primary platform up and operating, you’ll be able to be taught out of your customers’ suggestions and proper the course with out losing money and time. To take action, you could iterate and get to market rapidly. Specializing in a full-featured platform too early is counterproductive to this. As an alternative, prioritize transport core options for business-critical use circumstances.


The choice to construct or purchase an ML/AI platform is important and can affect your organization’s capability to ship for years to come back. By embarking on the method clear-eyed and thoroughly evaluating numerous elements, you may make an knowledgeable determination that may probably grow to be the best one.

As I’ve argued, a buy-and-build method represents one of the best technique. Following this philosophy, you’ll be able to both begin by constructing software program tailor-made to your fast wants or shopping for software program with minimal funding.

The important thing to long-term success isn’t to contemplate a platform “achieved.” As an alternative, be open to adapting to new inside processes, compliance wants, and the ever-changing landscape of MLOps tools.

Be happy to achieve out to me on LinkedIn, I’m very happy to reply any enter you might need.

Was the article helpful?

Thanks on your suggestions!

Thanks on your vote! It has been famous. | What subjects you want to see on your subsequent learn?

Thanks on your vote! It has been famous. | Tell us what must be improved.

Thanks! Your recommendations have been forwarded to our editors

Discover extra content material subjects:

Leave a Reply

Your email address will not be published. Required fields are marked *